DFT and DMFT: Implementations and Applications to the Study of Correlated Materials By ERIK RYAN YLVISAKER B.S. (Southern Oregon University) 2003 M.S. (University of California, Davis) 2004 DISSERTATION Submitted in partial satisfaction of the requirements for the degree of DOCTOR OF PHILOSOPHY in PHYSICS in the OFFICE OF GRADUATE STUDIES of the UNIVERSITY OF CALIFORNIA DAVIS Approved: Committee in Charge 2008 i
178
Embed
DFT and DMFT: Implementations and Applications …yclept.ucdavis.edu/Theses/Erik.thesis.pdfDFT and DMFT: Implementations and Applications to the Study of ... so that with small variations
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
DFT and DMFT: Implementations and Applications to the Study of Correlated Materials
By
ERIK RYAN YLVISAKERB.S. (Southern Oregon University) 2003
M.S. (University of California, Davis) 2004
DISSERTATION
Submitted in partial satisfaction of the requirements for the degree of
DOCTOR OF PHILOSOPHY
in
PHYSICS
in the
OFFICE OF GRADUATE STUDIES
of the
UNIVERSITY OF CALIFORNIA
DAVIS
Approved:
Committee in Charge
2008
i
Abstract
While DFT-LDA has enjoyed great success in describing many ground-state properties of
solids, there is an ever increasing list of materials which are not described even qualitatively
correct in DFT-LDA. Here I describe some applications of DFT and dynamical mean field
theory.
This dissertation is divided into two parts. Part I describes the theoretical background
of DFT and DMFT, and the simplest extension to DFT to study correlated materials,
LDA+U, is examined in detail. We find that the behavior of LDA+U can differ quite
strongly between AMF and FLL, the two commonly used double-counting functionals.
AMF has a strong energetic penalty for magnetic states, which roughly cancels the ex-
change splitting found in LSDA. In Part I, I also describe in detail the implementation of
LDA+DMFT in the publicly available code FPLO.
Part II focuses on applications. I describe the application of LDA to LiNbO2, where
Wannier functions and effective tight-binding Hamiltonians are constructed for LiNbO2.
We found that second neighbor hopping t2 is the largest, but the first neighbor hopping
depends strongly on the Nb-O distance, so that with small variations of O position t1 varies
by an order of magnitude. I also describe in part II the application of LDA in density
functional perturbation theory (DFPT) to calculate phonons for Al, Na and diamond to
compute melting curves using the Lindemann criteria. The resulting Tm(P ) curves agree
rather well with experiment in most conditions for these materials, including reproducing
the drop of 300 K of Tm in bcc-Na. Detailed calculations in LDA+DMFT using Hubbard
I and QMC impurity solvers are used to describe the valence transition in Yb. The
QMC calculations showed some unique ergodicity which proved difficult to deal with. The
agreement with experimental XAES and RIXS measurements of nf is rather good, and
even the highly approximate Hubbard I impurity solver gives reasonably good results.
Finally, I discuss the application of LDA+U to molecular orbitals in RbO2 to examine the
possibility of orbital ordering in the O π∗ orbitals. The orbital ordering in this system
is highly frustrated, which is likely responsible for the many low-temperature structural
ii
distortions seen experimentally. Two kinds of orbital ordering are found, depending on
the parameters U and U ′ which describe intraorbital repulsion and interorbital repulsion
In practice, it has been found that GGA functionals don’t perform significantly better
than LDA functionals in general. GGA usually leads to improvements in bond lengths and
lattice constants, but in some cases it leads to worse results. Common GGA functionals are
Perdew-Wang [13] and Perdew-Burke-Enzerhof (PBE) [2]. There exist several variations
on PBE, such as revPBE [14] and PBEsol [15] which are obtained by changing some of
the empirical parameters used in PBE.
CHAPTER 2. DENSITY FUNCTIONAL THEORY 12
2.4 Density Functional Perturbation Theory
Density functional perturbation theory (DFPT) [5] is a procedure developed for the cal-
culation of response functions. It is closely related to the method of linear response [4]. A
detailed description of the implementation of DFPT for phonons and electric field response
is given in references [16] and [17], and is quite lengthy. Practical applications of DFPT
have been used to calculate phonon frequencies and Born effective charges.
The general approach is to consider a perturbation to the external potential with the
expansion
vext = v(0)ext + λv
(1)ext + λ2v
(2)ext + λ3v
(3)ext + . . . , (2.26)
then other perturbed quantities E, H, ψkn, n(r), etc. are similarly expanded. Generally
the quantity of interest in applications of DFPT is the second order energy E(2), which
corresponds to taking two derivatives of the energy with respect to an external parameter
(or possibly two, if a second mixed partial of E is desired). This quantity is used in
calculation of the dynamical matrix for phonon calculations, and the computation of Born
effective charges. If one writes the density functional energy as
E[ψ] = minψ
∑
i∈occ
〈ψi|T + Vext|ψi〉 + EHxc[n] (2.27)
then the second order energy [4, 5, 16] is
E(2) = minψ(1)
∑
i∈occ
[
〈ψ(1)i |H(0) − ε
(0)i |ψ(1)〉 + 〈ψ(1)
i |v(1)ext|ψ
(0)i 〉 + 〈ψ(0)
i |v(1)ext|ψ(1)〉
+ 〈ψ(0)i |v(2)
ext|ψ(0)i 〉]
+1
2
∫∫
δ2EHxcδn(r)δn(r′)
∣
∣
∣
∣
n(0)
n(1)(r)n(1)(r′) d3r d3r′
+
∫
d
dλ
δEHxcδn(r)
∣
∣
∣
∣
n(0)
n(1)(r) d3r1
2
d2EHxcdλ2
∣
∣
∣
∣
n(0)
(2.28)
with the constraint that the first order wavefunctions are orthogonal to the ground state
CHAPTER 2. DENSITY FUNCTIONAL THEORY 13
wavefunctions,
〈ψ(0)i |ψ(1)
j 〉 = 0. (2.29)
The second order energy is variational with respect to the first order wavefunctions. Since
only the first order wavefunctions are needed to compute second derivatives of the energy,
one can compute second mixed partials of the energy with respect to different perturbations
using the wavefunction response to only one of the perturbations. This approach can be
used to calculate, eg. Born effective charges, where the second mixed partial of total energy
with respect to an electric field and atomic displacement is needed.
2.4.1 Phonons
For computation of phonons aside from the Γ point, one must consider the effect of non-
commensurate perturbations characterized by a wavevector q in the first Brillouin zone.
In this case, the ground state potential is periodic,
v(0)ext(r + R) = v
(0)ext(r), (2.30)
and the first order potential will satisfy
v(1)ext,q(r + R) = eiq·R v
(1)ext,q(r). (2.31)
This perturbation is non-Hermitian, so one must also consider the same perturbation with
wavevector −q. However, the response to the sum of perturbations will be just the sum
of the responses to the individual perturbation, so we can consider the non-Hermitian
operator for simplicity in this discussion. The first order responses in the wavefunction
and density satisfy
ψ(1)i,k,q(r + R) = ei(k+q)·R ψ
(1)i,k,q(r) (2.32)
n(1)q (r + R) = eiq·R n
(1)q (r). (2.33)
CHAPTER 2. DENSITY FUNCTIONAL THEORY 14
The perturbing potential for a collective atomic displacement where the κ-th atom
located at sκ is moved in the direction α will be given by
v(1)ext,q(r) =
∑
j
eiq·Rj∂
∂τκ,αvκ(r− Rj) (2.34)
v(2)ext,q(r) =
∑
j
1
2
∂2
∂2τκ,αvκ(r − Rj) (2.35)
where vκ is the potential from the atom being perturbed.
The calculation of phonon spectrum procedes as follows. First, for each atom in the
unit cell, one computes the dynamical matrix
Dκi,κ′j(R) =∂2E
∂sRκi∂s0κ′j(2.36)
by computing the response to atomic displacements for each atom in each direction. Then
the Fourier transformed dynamical matrix when diagonalized in the generalized eigenvalue
problem
D(q)s = Mω2q s (2.37)
yields the phonon frequencies at pseudomomentum vector q. Here, M is a matrix with the
atomic masses on the diagonal, and the resulting eigenvalues ω2q are actually the squares
of the phonon frequencies.
2.5 Basis Sets
In practice, an application of DFT requires that the Kohn-Sham eigenstates are expanded
in some finite basis set. The discussion up to this point in this chapter on DFT has
been general; there have been no boundary conditions assumed. In this section we will
discuss basis sets which are useful mainly for solids, so periodic boundary conditions will
be assumed.
There are a large variety of basis sets; a partial list of the most widely used ones would
CHAPTER 2. DENSITY FUNCTIONAL THEORY 15
include planewaves (PW), local orbitals (LO), linearized augmented planewaves (LAPW),
linearized muffin-tin orbitals (LMTO). In this section, I will discuss the two basis sets that
have been used most extensively in this research, which are PW and LO.
2.5.1 Planewaves
A basis set composed entirely of planewaves has the convenient property that with appro-
priate selection of planewaves, the basis automatically fulfills the periodicity condition of
a crystalline solid. The Kohn-Sham eigenstate ψnk is expanded in planewaves as
〈r|ψnk〉 =∑
G
cn,k+G ei(k+G)·r = eik·ruk(r) (2.38)
where r is a real-space vector, and G =∑
i nibi where ni are integers and the bi are the
reciprocal lattice vectors. The satisfaction of the Bloch condition is explicit; the function
uk(r) has the periodicity of the unit cell as it is composed of planewaves with wavevectors
that are integer multiples of the reciprocal lattice vectors.
For a practical implementation, the basis set needs to be cut off somewhere to contain
a finite number of planewaves. This is generally done with a single parameter Ecut which
is taken as the maximum kinetic energy of any planewave used in the expansion. This
results in the condition that planewaves are kept in the basis set expansion if they satisfy
the condition
1
2(k + G)2 ≤ Ecut (2.39)
Loosely speaking, this condition represents a sphere in reciprocal space inside which points
representing planewave vectors which satisfy the Bloch condition are kept in the basis
expansion, and planewaves with very high kinetic energy that fall outside the sphere are
discarded.
In practice, a planewave expansion is not practical to use for the basic solid problem
where the atomic cores have potential −Z/r and all electrons are treated on the same
CHAPTER 2. DENSITY FUNCTIONAL THEORY 16
footing. The difficulty with the planewave basis is caused by two factors. The first com-
plication is the treatment of highly localized core electrons. The wavefunctions for these
electrons is very rapidly varying in the region near the atomic cores, and rapidly goes
to zero away from the nucleus. Expanding this type of wavefunction would require a ex-
tremely large set of planewaves with extremely large kinetic energy to accurately represent
these core electrons. Such a large basis set would end up being completely unwieldy com-
putationally, and rather unnecessary as the core electrons do not play a role in many of
properties of interest in solids.
The second complication for planewaves can come in expanding the core region of
the valence wavefunctions, for electrons with a high principle quantum number. This is
because states with different quantum numbers must be orthogonal. For electrons with
the same angular momentum quantum numbers l and ml values but principle quantum
numbers n, the orthogonality cannot be taken care of by the angular part, so the radial
functions must have a different number of nodes to be orthogonal. As a general rule, for
hydrogenic wave functions the number of nodes in the radial wavefunction is n− l− 1, so
electrons with higher n will have extra nodes. These nodes always occur near the atomic
cores in solids. Correctly capturing these nodes in a planewave basis would require a very
large basis set.
The solution that is always used in planewave basis sets is to replace the atomic poten-
tial with a pseudopotential. There are a large variety of methods for generating pseudopo-
tentials, but all of the results presented in this dissertation using planewaves were done
with norm-conserving Troullier-Martins pseudopotentials [18]. These pseudopotentials are
non-local [19], and can optionally include a nonlinear core correction [20]. A non-local
pseudopotential contains different potentials for different orbital angular momentum val-
ues, and can be written
V =∑
lm
|lm〉Vl〈lm|. (2.40)
The basic approach to generating a Troullier-Martins pseudopotential is this. One
CHAPTER 2. DENSITY FUNCTIONAL THEORY 17
chooses which valence states should be used to generate the nonlocal potentials, and a
core-radius for each is chosen. Outside this radius, the pseudopotential is exactly the
atomic potential −Z/r, with Z equal to the number of valence electrons. Inside this
radius the pseudopotential becomes a smooth function which is not divergent at r = 0.
A detailed discussion of generating pseudopotentials, as well as pseudopotentials for all
elements up to Pu can be found in Reference [21].
2.5.2 Local Orbitals (FPLO)
Of all the common basis sets, planewaves could be (loosely) thought of as being the ‘de-
localized extreme,’ where each basis function is completely delocalized. The opposite
extreme would be a basis set composed entirely of localized, atomic-like orbitals. This is
what is implemented in the Full Potential Local Orbital (FPLO) code. [22]
The basis functions are computed by solving an atomic like problem with potential
vat(r) = − 1
4π
∫
v(r − R − s)dΩ +
(
r
r0
)m
. (2.41)
The first term in (2.41) is the spherical average of the crystal potential, and the second term
is the confining potential, with a typical value of m of 4. r0 is a compression parameter
which is recast in terms of a dimensionless parameter x0 such that r0 = (x0rNN/2)3/2 with
rNN the nearest neighbor distance, and then x0 is then taken as a variational parameter
in the self-consistent procedure. The basis functions that are used in the first step of the
calculation are computed just from the atomic potential vat(r) = −Z/r, but after the
density is computed in each DFT step, the basis functions are recalculated based on the
new crystal potential.
It is important to note that although the atomic orbitals are orthogonal, the basis
functions used in the crystal are not. The overlap matrix has form
S(k)ij = δij +
∑
R
〈Ri|0j〉eik·(R+si−sj) (2.42)
CHAPTER 2. DENSITY FUNCTIONAL THEORY 18
where si locates the i-th atom in the unit cell, and R is a Bravais lattice vector, so that
the vector R + si − sj goes between the i-th atom in the unit cell at the origin and the
j-th atom in the unit cell located at R. The overlap matrix is not the identity matrix
because the tail of a basis function on one atom overlaps the tail of basis functions on
other atoms. This changes the secular equation, requiring that a generalized eigenvalue
problem be solved,
|kn〉 =∑
j
cjkn∑
R
N−1/2R eik·R|Rj〉 (2.43)
∑
j′
[
H(k)jj′ − εknS
(k)jj′
]
cj′kn = 0 (2.44)
where |kn〉 is the wavefunction and cjkn is the coefficient in the basis expansion. The
second summation in (2.43) ensures that |kn〉 satisfies the Bloch theorem.
2.6 Wannier Functions
The use of Bloch waves as solutions in a crystalline solid allows us identify the pseudo-
momentum k as a good quantum number (or three, rather), and to block diagonalize the
full Hamiltonian so that there is only a series of k-dependent Hamiltonians. Visualiza-
tion of these wavefunctions is nontrivial however, because (1) Bloch waves are completely
delocalized and (2) they are k-dependent, so there is a very large number of different
wavefunctions to visualize, even for a single band.
A solution to both these problems is available by constructing of Wannier functions.
Wannier functions are essentially a Fourier transform of the wavefunctions labeled by k to
basis functions which exist at each site in the lattice. Additionally, the Wannier functions
are independent of R, so that the same Wannier functions sit at every lattice site. In
this way, one needs to visualize no more than one Wannier function per electronic band
of interest. Wannier functionals also incorporate the effects of bonding that occur in the
crystal, so for this reason they are sometimes colloquially referred to as ‘molecular orbitals
CHAPTER 2. DENSITY FUNCTIONAL THEORY 19
of solids.’
The construction of Wannier functions is as follows. Them-th Wannier function located
at site R, designated as |Rm〉 is given by
|Rm〉 = N−1/2k
∑
k
Ukmne
−ik·R|kn〉 (2.45)
where Uk is a unitary matrix. Uk need not be a square matrix actually, this procedure
can be applied by projecting out M Wannier functions from N bands where M ≤ N . In
this case, the rows of Uk must represent a set of orthogonal vectors.
The appearance of Uk in (2.45) allows a gauge freedom in this transformation. In
practice, the wavefunctions |kn〉 come from diagonalizing the Hamiltonian matrix in some
basis, and of course any eigenvector returned by a numerical diagonalization routine may
come with an attached phase factor eiα. Numerical diagonalization routines don’t generally
make any guarantees about what phase factors are attached to the eigenvectors, so it may
not even be consistent or continuous at different k-points. The inclusion of Uk in the sum
(2.45) allows one to normalize the phase factors for each wavefunction included in the sum
so that the wavefunctions constructively interfere near the desired site, and destructively
interfere away from the site so that the Wannier function is localized. There are a few
different approaches to choosing the Uk matrices, such as Maximally Localized Wannier
Functions [23, 24], or by using projections onto local orbitals. [25, 26]
The existence of the gauge freedom allows for a certain degree of arbitrariness in the
resultant Wannier functions. It is generally desirable to have Wannier functions that are
both localized and real, which greatly cuts down on the gauge freedom. It seems that in
every example considered in the literature both these properties come together, however
there is no general proof that this is true in three dimensions or with multiple entangled
bands. Such a proof is only known for a single band in one dimension. [27]
CHAPTER 2. DENSITY FUNCTIONAL THEORY 20
2.7 Tight Binding
The tight binding model allows one to construct wavefunctions for a periodic crystal from
localized atomic orbitals. It is sometimes called linear combination of atomic orbitals
(LCAO). Since the tight binding approach is detailed in many introductory solid-state
physics texts [8, 28, 29], I will only present a brief overview of tight-binding here as it
applies to the applications discussed in chapters 6 and 9.
The goal of the tight binding approach is to construct a representation of the Hamil-
tonian and its wavefunctions in a basis of localized atomic orbitals |Rb〉, which represents
the b-th atomic orbital at site R. The real space projection is 〈r|Rb〉 = φb(r − R) where
φb(r) is the real space atomic orbital. Henceforth we shall assume that the on-site orbitals
are orthogonal, 〈Rb|Rb′〉 = δbb′ .
The first thing to do is construct basis states that satisfy the Bloch condition, by
defining
|kb〉 = N−1/2∑
R
eik·R|Rb〉. (2.46)
Then the Hamiltonian can be calculated via the overlap
Hkb,b′ ≡ 〈kb|H|kb′〉 = εb δb,b′ +
∑
R 6=0
tb,b′(R)eik·R (2.47)
where εb is the on-site energy 〈0b|H|0b〉 and tb,b′(R) = 〈0b|H|Rb′〉 is the tight binding
coefficient, often interpreted as the amplitude for hopping from an orbital b at the origin
to an orbital b′ at a distance R.
This basis has not been constructed in a manner that implies orthogonality. The atomic
orbitals used in the basis are not (necessarily) orthogonal to orbitals residing on different
sites. Thus, it is necessary to calculate the overlap matrix
Skb,b′ ≡ 〈kb|kb′〉 = δb,b′ +
∑
R 6=0
sb,b′(R)eik·R (2.48)
CHAPTER 2. DENSITY FUNCTIONAL THEORY 21
where sb,b′(R) = 〈0b|Rb′〉 is the overlap integral between orbital b on the site at the origin
with another orbital b′ at a site R.
The appearance of a non-identity overlap matrix necessitates that the secular equation
to be solved takes the form of a generalized eigenvalue problem,
Hk|kn〉 = εknSk|kn〉. (2.49)
If one chooses Wannier functions as defined in section 2.6 for the atomic orbitals, then the
overlap matrix (2.48) becomes the identity matrix and one can solve the regular eigenvalue
problem.
What is often done in tight-binding fits is to take a bandstructure εkn from an LDA
calculation and treat the tight-binding parameters εb and tb,b′(R) as fitting parameters,
along with the assumption that the local orbitals are Wannier functions so that the overlap
matrix is the identity matrix. This works well for systems with an isolated set of bands
without large bandwidths. In this case it is generally expected that one does not need many
parameters to get a good fit to the DFT band structure. The tight binding parameters
can then be used in model Hamiltonian calculations for more sophisticated many-body
methods. For more detailed examples of tight-binding fits, see chapters 6 and 9.
22
Chapter 3
LSDA+U
The work described in this chapter was done in collaboration with Warren E. Pickett and
Klaus Koepernik.
3.1 Introduction
Density functional theory (DFT) and its associated local (spin) density approximation
[L(S)DA] is used widely to describe the properties of a wide variety of materials, often
with great success. However there exists a class of materials which are poorly described,
sometimes qualitatively, by LDA. These so-called strongly correlated materials typically
contain atoms with open d or f shells, in which the corresponding orbitals are in some sense
localized. The LSDA+U approach was introduced by Anisimov, Zaanen, and Andersen[30]
to treat correlated materials as a modification of LDA (‘on top of LDA’) that adds an intra-
atomic Hubbard U repulsion term in the energy functional. Treated in a self-consistent
mean field (‘Hartree-Fock’) manner, in quite a large number of cases the LDA+U result
provides a greatly improved description of strongly correlated materials.
At the most basic level, the LSDA+U correction tends to drive the correlated orbital
m occupation numbers nmσ (σ denotes spin projection) to integer values 0 or 1. This in
turn produces, under appropriate conditions, insulating states out of conducting LSDA
CHAPTER 3. LSDA+U 23
states, and the Mott insulating state of several systems is regarded as being well described
by LSDA+U at the band theory level. Dudarev et al.[31] and Petukhov et al.[32] provided
some description of the effect of the spin dependence of two different double counting
terms within an isotropic approximation. Beyond this important but simple effect, there
is freedom in which of the spin-orbitals (mσ) will be occupied, which can affect the result
considerably and therefore makes it important to understand the effects of anisotropy and
spin polarization in LSDA+U. After the successes of providing realistic pictures of the
Mott insulating state in La2CuO4 and the transition metal monoxides,[30] the anisotropy
contained in the LSDA+U method produced the correct orbitally ordered magnetic ar-
rangement for KCuF3 that provided an understanding of its magnetic behavior.[33]
The anisotropy of the interaction, and its connection to the level of spin polarization, is
a topic that is gaining interest and importance. One example is in the LSDA+U descrip-
tion of the zero temperature Mott transition under pressure in the classic Mott insulator
MnO. The first transition under pressure is predicted to be[34] an insulator-insulator (not
insulator-metal) transition, with a S = 52 → S = 1
2 moment collapse and a volume collapse.
The insulator-to-insulator aspect is surprising, but more surprising is the form of moment
collapse: each orbital remains singly occupied beyond the transition, but the spins of elec-
trons in two of the orbitals have flipped direction. This type of moment collapse is totally
unanticipated (and hence disbelieved by some), but it is robust against crystal structure
(occurring in both rocksalt and NiAs structures) and against reasonable variation of the
interaction strength. Detailed analysis indicates it is a product of the anisotropy of the
LSDA+U interaction and the symmetry lowering due to antiferromagnetic order.
Another unanticipated result was obtained[35] in LaNiO2, which is a metal experimen-
tally. This compound is also a metal in LSDA+U over a very large range of interaction
strength U , rather than reverting to a Mott insulating Ni1+ system which would be iso-
valent with CaCuO2. For values of U in the range expected to be appropriate for the
Ni ion in this oxide, the magnetic system consists of an atomic singlet consisting of an-
tialigned dx2−y2 and dz2 spins on each Ni ion. Again the anisotropy of the interaction
CHAPTER 3. LSDA+U 24
evidently plays a crucial role in the result, with its effect being coupled thoroughly with
band mixing effects.
The addition of a Hubbard U interaction also introduces the need for “double counting”
correction terms in the energy functional, to account for the fact that the Coulomb energy
is already included (albeit more approximately) in the LSDA functional. All double count-
ing schemes subtract an averaged energy for the occupation of a selected reference state
depending only on Nσ, which largely cancels the isotropic interaction of the EI term
Eq. (3.2). Several forms for these double-counting terms have been proposed,[30, 36, 37]
but primarily two are commonly used. These LDA+U functionals are most often referred
to as around mean field (AMF) and the fully localized limit (FLL), which is also referred
to as the atomic limit (AL). The distinctions between these forms have attracted some
discussion, but without consideration of the full anisotropy of the interaction.
The need for double-counting corrections is not unique to the LDA+U method; any
other method that adds correlation terms to the LSDA functional, such as the dynamical
LDA+DMFT (dynamical mean field theory) approach, also will require double-counting
corrections. This is an unfortunate consequence of LDA’s success; LDA works too well,
even in correlated systems where it usually gets interatomic charge balance reasonably, to
just throw it away.1 The common approach has been to use LSDA for correlated materials
and include a double-counting correction. There are techniques being developed which do
not build on a correction to DFT-LDA, but it remains to be seen whether these approaches
will be successfully applied to a broad range of solid-state materials.
Although there has been much study on the performance of these LDA+U functionals
in the context of real materials, and an early review of the method and some applications
was provided by Anisimov, Aryasetiawan, and Lichtenstein,[38] relatively little has been
done to understand, qualitatively and semi-quantitatively, how the functionals operate
based solely on their energetics, distinct from DFT-LSDA effects. In this paper we analyze
1There are numerous examples for strongly correlated (heavy fermion) metals where the Fermi surfacecalculated within LDA is predicted as well as for more conventional metals. This surely requires that thecharge balance between the various atoms is accurate.
CHAPTER 3. LSDA+U 25
the functionals that are commonly used, as well as others which were introduced early on
but are not so commonly used. Some of the nomenclature in the literature is confusing,
so we try to clarify these confusions where we can.
CH
AP
TE
R3.LSD
A+
U26
LDA+U Edc = Edc DFT XCFunctional (rewritten) Functional
Fl-nS 12UN
2 − U+2lJ2l+1
14N
2 = 12UN
2 − U+2lJ2l+1
12
∑
σ(N2 )2 LDA
Fl-S (AMF) 12UN
2 − U+2lJ2l+1
12
∑
σN2σ = 1
2UN2 − U+2lJ
2l+112
∑
σ N2σ LSDA
FLL 12UN(N − 1) − 1
2J∑
σNσ(Nσ − 1) = 12UN(N − 1) − 1
2J∑
σ(N2σ −Nσ) LSDA
FLL-nS 12UN(N − 1) − 1
4JN(N − 2) = 12UN(N − 1) − 1
2J∑
σ((N2 )2 −Nσ) LDA
Table 3.1: The double-counting terms of various LDA+U functionals. In the second expression two of them are rewritten toreflect how they are (somewhat deceptively) identical in form, but in one case a distinction between spin-up and spin-down(relative to half of N: Nσ ↔ N/2) is made. Note that while the first two forms appear to contain an isotropic self-interaction[12UN
2 rather than 12N(N − 1)] they are derived from a form which explicitly has no self-interaction between the orbital
fluctuations δMσ . See text for more discussion.
CHAPTER 3. LSDA+U 27
3.2 The LSDA+U Correction ∆E
The LDA+U functional is usually coded in a form in which the choice of coordinate sys-
tem is irrelevant, often referred to as the rotationally-invariant form.[33] This form involves
Coulomb matrix elements that have four orbital indices, and the orbital occupation num-
bers are matrices in orbital space (viz. nmm′). One can always (after the fact) rotate into
the orbital Hilbert space in which the occupations are diagonal, in which case the interac-
tions have only two indices. In our discussion we will work in the diagonal representation.
The LDA+U functionals considered here can all be written in the form
∆E = EI − Edc, (3.1)
where the direct interaction is
EI =1
2
∑
mσ 6=m′σ′
W σσ′mm′nmσnm′σ′ (3.2)
and Edc is the double-counting correction. The Coulomb matrix elements are given in
terms of the direct and (spin-dependent) exchange contributions as
W σσ′mm′ = (Umm′ − Jmm′δσ,σ′). (3.3)
By the convention chosen here, EI and Edc are both positive quantities as long as the
constants U and J (which define the matrix elements Umm′ and Jmm′ but are not the
same) as chosen conventionally, with U much larger than J .
Note that the orbital+spin diagonal term has been omitted in Eq. 3.2 – there is
no self-interaction in EI . However, it is formally allowed to include the diagonal ‘self-
interaction’ term, because the matrix element vanishes identically (self-interaction equals
self-exchange: Umm = Jmm), and it can simplify expressions (sometimes at a cost in
clarity) if this is done. The double counting correction depends only on the orbital sum
CHAPTER 3. LSDA+U 28
Nσ, which appears up to quadratic order. A consequence is that it will contain terms in
nmσnmσ, which are self-interactions. Thus while the LSDA+U method was not intended
as a self-interaction correction method, it is not totally self-interaction free. In fact, the
underlying LSDA method also contains self-interaction, and the double-counting term may
serve to compensate somewhat this unwanted effect. We discuss self-interaction at selected
points in this paper.
3.2.1 Short formal background to the LSDA+U method.
The “LSDA+U method” is actually a class of functionals. Each functional has the same
form of interaction EI , with differences specified by
(1) choice of the form of double counting term.
(2) choice of constants U and J . For a given functional, these are ‘universal’ constants like
~,m, e, i.e. they are not functional of the density in current implementations. Possibilities
for doing so, that is, determining them self-consistently within the theory, have been
proposed.[39]
(3) choice of projection method to determine the occupation matrices from the Kohn-Sham
orbitals. Given identical choices for (1) and (2) above, there will be some (typically small)
differences in results from different codes due to the projection method.
The occupation numbers (or, more generally, matrices) are functionals of the density,
nmσ[ρ], through their dependence on the Kohn-Sham orbitals. Then, whereas in LSDA
one uses the functional derivative
LSDA :∂ELSDA[ρs]
∂ρσ(r)(3.4)
in minimizing the functional, in LSDA+U the expression generalizes to
LSDA + U :∂(ELSDA[ρs] + ∆E[nms[ρs]])
∂ρσ(r)(3.5)
CHAPTER 3. LSDA+U 29
Since the resulting spin densities ρs are changed by including the ∆E correction, the change
in energy involves not only ∆E but also the change in ELSDA. In practice, there is no
reason to compare ELSDA+U with ELSDA as they are such different functionals. However,
in the following we will be assessing the importance of the choice of the double counting
term in the LSDA+U functional, and it is of interest to compare, for fixed U and J , the
energy differences between LSDA+U functionals differing only in their double counting
terms in order to understand the differing results. Even if the set of occupation numbers
turn out to be the same (a situation we consider below), the densities ρσ will be different
and the differences in ELSDA may become important.
As with the non-kinetic energy terms in ELSDA, the functional derivatives of ∆E lead
to potentials in the Kohn-Sham equation. These are non-local potentials, which (via the
same projection used to define the occupation numbers) give rise to orbital-dependent
(nonlocal) potentials
vmσ ≡ ∂∆E
∂nmσ= vImσ − vdcmσ,
vImσ =∑
m′σ′ 6=mσ
W σσ′mm′nm′σ′ . (3.6)
The double counting orbital potential is discussed later.
The corresponding contribution to the eigenvalue sum Esum is
∆Esum =∑
mσ
vmσnmσ, (3.7)
which is subtracted from the eigenvalue sum to obtain the Kohn-Sham kinetic energy.
However, there are indirect effects of the orbital potentials that affect all of the kinetic
and (LSDA) potential energies; these will be different for different ∆E functionals because
the orbital potentials, which depend on the derivative of ∆E and not simply on the values
of nmσ, differ for each functional. This makes it necessary, for understanding the effects of
the ∆E correction and the change in energy, to analyze the orbital potentials. We provide
CHAPTER 3. LSDA+U 30
a brief discussion in Sec. V.
3.2.2 Fluctuation forms of LSDA+U
First we consider the class of functionals that can be written in what is termed here as a
fluctuation form. The original LDA+U functional was introduced in 1991 by Anisimov,
Zaanen and Andersen[30] and was written as
∆EF l−nS =1
2
∑
mσ 6=m′σ′
W σσ′mm′(nmσ − n)(nm′σ′ − n), (3.8)
where n = N corr/2(2l+1) is the average occupation of the correlated orbitals. (Henceforth
N ≡ N corr.) Note that the energy is changed only according to angular ‘fluctuation’ away
from the (spin-independent) angular average occupation. This form is properly used with
LDA (the ‘LDA averages’ n are the reference) and not LSDA. This form was originally
advocated with generic (U − Jδσ,σ′ ) matrix elements instead of the full Coulomb matrix,
but we use the full W σσ′mm′ here for comparison with other functionals.
In 1994, Czyzyk and Sawatzky[37] introduced a change to (3.8) and also proposed a new
functional. The motivation for changing (3.8) was to use an LSDA exchange-correlation
functional to treat spin splitting effects rather than LDA. This change motivated the
following equation,
∆EF l−S =1
2
∑
mσ 6=m′σ′
W σσ′mm(nmσ − nσ)(nm′σ′ − nσ′)
= ∆EAMF (3.9)
where nσ = Nσ/(2l+1) is the average occupation of a single spin of the correlated orbitals.
Here the energy correction is due to angular fluctuations away from the spin-dependent
angular mean, and hence must be used with LSDA. We point out that the authors in Ref.
[[37]] refer to Eq. (3.8) as ELDA+AMF and Eq. (3.9) as ELSDA+AMF. This wording may
have caused subsequent confusion, due to the way these terms have come to be used, and
CHAPTER 3. LSDA+U 31
also because a discussion of the “+U” functionals requires explicit specification of whether
LDA or LSDA is being used just to understand which functional is being discussed. Also
confusing is that Solovyev, Dederichs, and Anisimov[40] rejustified Eq. 3.8 using “atomic
limit” terminology.
The fluctuation forms of LSDA+U are automatically particle-hole symmetric, since
nmσ → 1 − nmσ, nσ → 1 − n gives nmσ − nσ → −(nmσ − nσ) and the expression is
quadratic in these fluctuations. The general form of Eq. 3.1 need not be particle-hole
symmetric.
Many authors (present authors included) have used the term LDA+U where the term
LSDA+U would be more appropriate, which is especially confusing when discussing the
AMF functional. We choose to depart from this confusing nomenclature by giving (3.8)
and (3.9) unique names specifying their fluctuation forms, and their connection to LSDA
(Fl-S) or to LDA (Fl-nS). We collect the double-counting terms for the various functionals,
along with their connection to LDA or LSDA, in Table I.
3.2.3 The Fully Localized Limit (FLL) Functional
The second functional introduced by Czyzyk and Sawatzky[37] is the FLL functional. (A
J=0 version of FLL was introduced in 1993 by Anisimov et al.[36]) The authors referred
to it (confusingly, as terminology has progressed) as the “around mean field” functional
but the atomic limit double counting term. This double-couting term can be obtained by
considering the energy of an isolated atomic shell with N electrons, so in the literature it
is commonly referred to as the atomic limit, or fully localized limit (FLL) functional. This
functional cannot be written in the fluctuation form of the previous two functionals (the
fluctuation form is exhausted by the -S and -nS cases). The FLL functional is written in
the form of (3.1), with the double-counting term given in Table 3.1.
There is yet another LDA+U functional that is available, which was introduced in 1993
by Anisimov et al. [36] There is no clear name for it, but since it can be obtained by using
Nσ = N/2 in Edc for FLL, one might consistently refer to it as FLL-nS, corresponding to
CHAPTER 3. LSDA+U 32
FLL with no spin dependence. The authors in Ref. [36] indicate that this functional is to
be used with LDA, in accordance with the lack of spin dependence in the double counting
term.
3.2.4 Implementation of LSDA+U in Some Widely Used Codes
The Fl-nS functional is implemented in the Wien2k code, as nldau = 2, and called HMF
(Hubbard in Mean Field),[41] however it is apparently not often used.
The Fl-S (AMF) functional is implemented in the Wien2k code[41] as nldau = 0 and the
FPLO code[22] as AMF.
The FLL functional is implemented in several general-purpose DFT codes, such as Wien2k
(nldau = 1)[41], FPLO (select AL in fedit)[22], VASP and PW/SCF.
The FLL-nS functional is available in VASP.
3.2.5 General Remarks
When the Fl-nS and Fl-S functionals are written in their fluctuation form, there is no sep-
arate double-counting term, hence one does not need the double-counting interpretation.
They can of course be expanded to be written in the ‘interaction minus double-counting’
form of (3.2), which is useful especially for comparing with functionals that can only be
written in that form. A comparison of the double-counting terms is given in Table 3.1.
Reducing all to interaction minus double-counting form makes the difference between the
functionals most evident; since they all have the same “direct-interaction” term, the only
difference between the functionals is what double-counting energy is used; the uninteresting
tail seems to be wagging the exciting dog, which is in fact the case. The double-counting
terms can be reduced to dependence only on N and Nσ thanks to summation rules (there
CHAPTER 3. LSDA+U 33
is at least one free index) on the Umm′ and Jmm′ matrices,
∑
m
Umm′ = (2l + 1)U (3.10)
∑
m
Jmm′ = U + (2l)J, (3.11)
that is, the sum over any column (or row) of the U and J matrices is a fixed simple value,
which depends on the input parameters U and J . One can then simply see that a sum
over a column of W is (2l + 1)U if σ 6= σ′ and 2l(U − J) if σ = σ′.
The Umm′ and Jmm′ matrices satisfy, by definition, Umm−Jmm = 0, so that there is no
self-interaction, whether or not the (vanishing)diagonal term mσ = m′σ′ is included in the
interaction term. As mentioned earlier, the following analysis assumes the the occupation
matrix has been diagonalized. While this can always be done, the transformed matrix
elements Umm′ and Jmm′ will not be exactly what we have used in Sec. 3.6.
3.3 Analysis of the Functionals
The J = 0 simplification.
It is not uncommon for practitioners to use ‘effective’ values U = U − J, J = 0 and insert
these constants (for U, J) into LDSA+U. For J = 0, of course Hund’s coupling (intra-
atomic exchange) is lost, but J also controls the anisotropy of the interaction, and for
J = 0 anisotropy also is lost (Umm′ ≡ U as well as Jmm′ ≡ 0 for m 6= m′). This case is
relatively simple, it seems it should provide the “big picture” of what LSDA+U does with
simple Coulomb repulsion, and it has been discussed several times before. With J = 0,
CHAPTER 3. LSDA+U 34
the fluctuation functionals simplify to
∆EF l−κJ=0 =U
2
∑
mσ 6=m′σ′
δnmσδnm′σ′
=U
2
(
∑
mσ
δnmσ
)2
−∑
mσ
(δnmσ)2
= −U2
∑
mσ
(δnmσ)2 ≡ −U
2Γ2κ ≤ 0, (3.12)
because the sum of fluctuations vanishes by definition for either form κ = nS or S; note
the ‘sign change’ of this expression when the diagonal terms are added, and subtracted,
to simplify the expression. Here Γ2 is the sum of the squares of the fluctuations, bounded
by 0 ≤ Γ2κ ≤ N . For integer occupations the energy corrections for Fl-nS and Fl-S (AMF)
can be written
∆EF l−SJ=0 = −U2
[
N(1 − n) − M2
2(2l + 1)
]
,
∆EF l−nSJ=0 = −U2N(1 − n). (3.13)
There are two things to note here.
1. In Fl-nS, the energy is independent of both the spin and orbital polarization of the
state, which lacks the basic objective of what LSDA+U is intended to model. Con-
sidering the form of its double counting term (see Table 3.1) with its self-interaction
term (proportionality to N2), Fl-nS for J=0 becomes simply a self-interaction cor-
rection method.
2. In Fl-S (AMF), configurations with magnetic moments are energetically penalized,
proportionally to U and quadratically with M . In later sections we will discuss the
partial cancellation with the LSDA magnetic energy.
CHAPTER 3. LSDA+U 35
Under the same conditions, the FLL functional becomes
∆EFLL =U
2
∑
mσ
nmσ(1 − nmσ) ≥ 0. (3.14)
Solovyev et al.[40] noted the important and easily recognizable characteristics of this ex-
pression. Besides being non-negative, for integer occupations the energy vanishes. It is a
simple inverted parabola as a function of each nmσ. From the derivative, the orbital po-
tentials are linear functions of nmσ, with a discontinuity of U when nmσ crosses an integer
value. These characteristics underlie the most basic properties of the LSDA+U method:
integer occupations are energetically preferred, and discontinuities in the potentials model
realistically the Mott insulator gap that occurs in strongly interacting systems at (and
only at) integer filling.
J 6= 0, but Isotropic
Simplification of the full expression for a functional results by separating out the isotropic
parts of the interaction:
Umm′ = U + ∆Umm′ (3.15a)
Jmm′ = Uδmm′ + J(1 − δmm′) + ∆Jmm′ . (3.15b)
The isotropic parts simplify, giving
∆EF l−nS = −U − J
2
∑
mσ
n2mσ −
J
4M2
+U − J
2Nn+ ∆Eaniso, (3.16)
∆EF l−S = −U − J
2
∑
mσ
n2mσ +
U − J
4
M2
2l + 1
+U − J
2Nn+ ∆Eaniso, (3.17)
∆EFLL = −U − J
2
∑
mσ
n2mσ +
U − J
2N
+∆Eaniso (3.18)
CHAPTER 3. LSDA+U 36
with the universal anisotropy contribution
∆Eaniso =1
2
∑
mm′σσ′
∆W σσ′mm′nmσnm′σ′ , (3.19)
∆W σσ′mm′ ≡ ∆Umm′ − ∆Jmm′δσσ′ . (3.20)
is the anisotropic part of the interaction matrix elements. These equations, up to the ∆W
term, are the extensions of Eq. (3.13) to include isotropic exchange in explicit form.
The first term in each of these expressions contains −12 Un
2mσ(U ≡ U−J) and hence has
the appearance of a self-interaction correction. Since the diagonal term of the interaction
EI is specifically excluded, it does not actually contain any self-interaction; in fact, the sign
of the interaction EI is positive. (The double counting term does contain terms quadratic
in N which must be interpreted as self-interaction.) Nevertheless, the rewriting of the
functional leads to a self-interaction-like form, and that part of the functional will have
an effect related to what appears in the self-interaction-corrected LDA method, but by an
amount proportional to U rather than a direct Coulomb integral, and depending on the
difference of nmσ from the reference occupation, see Sec. IV.
Fl-nS
For Fl-nS, if we are restricted to integer occupations (so n2mσ = nmσ), then Γ2 depends
only on N , so the first term in ∆EF l−nS above depends only on N . Then, up to corrections
in ∆U and ∆J , the state with the largest total spin moment will be favored; this is Hund’s
first rule. In fact, even with the ∆U and ∆J terms, the −JM2/4 term is still strongly
dominant. Except for N = 7, there are many ways to arrange electrons in orbitals which
maximizes S. Energy differences between these arrangements arise only from anisotropy
(∆U and ∆J) and spin-orbit coupling.
CHAPTER 3. LSDA+U 37
Fl-S
In Fl-S, instead of having the −JM2/4 term from Fl-nS which favors magnetism, there
is a term (U−J)4(2l+1)M
2 which opposes magnetism. This term (as in the J = 0 case) comes
from the occupation variance which wants to evenly distribute electrons across both spin
channels. Within LSDA there is something like a Stoner term of the form −14IM
2 which
will compete with this Fl-S magnetic penalty. We return to this aspect in later sections
and the appendix.
Spin-orbit Coupling; Particle-Hole Symmetry
Without spin-orbit interaction, for a given N there are many states that are degenerate
for both double counting schemes. Every value of N has at least four degeneracies, those
with ±Lz,±Sz.
Any state which has the same number of spin up as spin down electrons (M = 0) gives
the same energy from Fl-nS and Fl-S, since then n↑ = n↓ = n (the orbital potentials are
distinct, however). Of course this fixed N , M=0 specification may contain many different
configurations. Looking at results mentioned later, for Fl-S the ground state for an even
number of electrons is Sz = 0 (so nσ = n), thus the configuration which gives the Fl-S
ground state has the same energy in Fl-S and Fl-nS.
3.4 Fractional Occupations
Here we briefly discuss the effect of non-integer occupations in LSDA+U. Taking a gen-
eral set of occupations as nmσ, we define a set of integer occupations, nmσ, and the
fractional part of the occupations as γmσ = nmσ − nmσ. For illustration purposes we will
choose the simplest possible scenario, where charge is transfered to an empty orbital a
from an occupied orbital b both of the same spin, so that 0 < γa↑ = −γb↑, na↑ = 0 and
nb↑ = 1. With this selection, Nσ is unchanged (and therefore, N and M as well) so that
Edc is unchanged. Thus, the effect of the charge transfer is entirely contained in the EI
CHAPTER 3. LSDA+U 38
term. Expanding EI for the general occupation set gives
EI [nmσ] − EI [nmσ] = Uγa↑(1 − γa↑) (3.21)
for the J = 0 case, and for J 6= 0 we find
EI [nmσ] − EI [nmσ]
=∑
mσ
(W ↑σam −W ↑σ
bm)nmσγa↑
−W ↑↑ab γ
2a↑. (3.22)
The dominant term in (3.22) is where mσ = b ↑. This term gives a contribution
W ↑σ′
ab γa↑ ∼ Uγa↑ (since U >> J for typical parameter choices, where other terms give
contributions proportional to (W ↑σam −W ↑σ
bm)γa↑ ∝ Jγa↑. The term with mσ = a ↑ is killed
off by the factor of na↑, and the term in γ2 is significantly smaller than the others for
γ < 0.5.
This shows that there is an energy penalty for fractional occupation, proportional to
U and linear in γ at small γ. Thus, in configuration space, the LSDA+U functionals have
many local minima around configurations with integer occupations. This result is fairly
general. Even for charge transfer between orbitals of opposite spins, the linear energy
penalty in γ is still dominant over any additional terms coming from the double-counting
or spin-orbit.
In practice, this gives the possibility that LSDA+U will get ‘stuck’ in a local minimum
with some configuration that may not be the true ground state. This behavior is not
uncommon; LSDA+U has been reported[42] to find multiple local minima depending on
the starting configuration.
CHAPTER 3. LSDA+U 39
3.5 Orbital Potential Matrix Elements
Up to now only the energy functionals themselves were discussed. Now we return to the
derivatives, the orbital potentials vmσ . It is simple to derive the exact expressions, and
the interaction term EI common to all forms gives a potential ∆vmσ which depends only
on the occupations of the other orbitals nm′σ′ ,m′σ′ 6= mσ. The potential resulting from
the double counting term is functional specific, and may contain a contribution from nmσ
itself, ı.e. a self-interaction.
We confine our observations here to the subdivision (introduced just above) of the
interaction into an unitarily invariant isotropic part, and into an anisotropic part Eq.
(3.15) that is much smaller and more difficult to analyze. As for the energy itself, it is
convenient to add and subtract the diagonal self-Coulomb and self-exchange, which makes
the effect of the potential much more transparent at the cost of introducing the misleading
self-interaction interpretation.
The potential matrix elements are
∆vF l−nSmσ = − (U − J) [nmσ − n] − J
2Mσ
+∆vanisomσ , (3.23)
∆vF l−Smσ = − (U − J) [nmσ − nσ] +U − J
2
M
2l + 1σ
+∆vanisomσ , (3.24)
∆vFLLmσ = − (U − J)
[
nmσ −1
2
]
+ ∆vanisomσ , (3.25)
with the anisotropic potential term
∆vanisomσ =
∑
m′σ′
∆W σσ′mm′nm′σ′ . (3.26)
The main occupation number dependent term, proportional to nmσ, has a self-interaction
appearance and effect, as discussed above for the functionals. The differences in this
CHAPTER 3. LSDA+U 40
term arise from the “reference” occupation with which nmσ is compared to determine the
potential shift. The “fluctuation” nmσ−nref is smaller for Fl−S (AMF) than for Fl−nS
because the occupation for a given spin direction tends to be closer to nσ than to n. The
reference occupation for FLL is, like Fl − nS, spin-independent, in fact, the reference is
half-filling. In this sense, FLL seems more like a single-band Hubbard model treatment
than the other two functionals.
The other difference that is evident in this form is the spin dependence. Fl − nS
additionally has a spin orientation dependent potential shift proportional to J and to
M (similar to an LSDA treatment, but using J instead of the Stoner I) and enhances
spin-splitting of the eigenenergies ε accordingly. In Fl-S (AMF) the analogous term is
+ (U − J) M2(2l+1)σ, with a sign that impedes magnetism. It can be simplified to ≈ J
2Mσ
when U ≈ 2 (l + 1) J . This expression illuminates the reason that AFM is sometimes found
to decrease the magnetic moment: this term more or less cancels the spin splitting of LSDA
due to the opposite sign. What is left is a splitting of occupied and unoccupied levels due to
the nmσ term, which is almost independent of M . The effect is to support a spin-polarized
solution, but provide little discrimination between different M . Since the spin polarization
energy does not favor large M , we end up with a tendency of a near degeneracy of different
M values, as we already pointed out from purely energetic arguments. For the case of a
half-filled fully polarized shell nms = δσ,1 (the case N = M = 7 in Section 3.6) the potential
matrix vanishes, which can be seen from n = 12
M2(2l+1) = n = 1
2 . However, at the same
time the energy contribution also vanishes ∆EF l−S = 0 (for integer occupations) and the
Fl-S functional has no effect at all.
The SIC term in FLL splits occupied and unoccupied states symmetrically, while in
the fluctuation functionals the splitting happens with respect to the averaged occupation,
which is seen in the overall energy positions in Fig 3.3.
CHAPTER 3. LSDA+U 41
0 2 4 6 8 10 12N
0
1
2
3
4
5
6
7
8
Ang
ular
Mom
entu
m
SLJ
0 2 4 6 8 10 12 14
(a) (b)
Figure 3.1: (Color online) Angular momentum values of Sz, Lz, and Jz of the lowestenergy state for (a) AMF (Fl-S) and (b) FLL, with spin-orbit coupling. Parameter valuesare U = 8, J = 1, I = 0.75. The AMF (Fl-S) curves do not follow Hund’s rules, becausethe Stoner parameter is too small. FLL follows Hund’s rules exactly with these parameters.
3.6 Numerical Results
Following common terminology, for the remainder of the paper we refer to the Fl-S func-
tional simply as the AMF form. We have taken values for Umm′ and Jmm′ (used for Eu)
from Ref. [43] (recalculated, to include more significant figures). These matrices are gen-
erated using U = 8 and J = 1 (values typical of rare earths) following the procedure given
in the appendix of Ref. [37].
In our analysis of the AMF and FLL functionals, which are based on an LSDA reference
state, we include a Stoner term
E(M) = −1
4IM2 (3.27)
to model the magnetic effects of LSDA on the total energy. The addition of this term helps
to give a picture of the degree to which the functionals reproduce Hund’s first rule. Typical
values of I for ionized lanthanides are 0.75 eV, so we use this value for the calculations of
CHAPTER 3. LSDA+U 42
this section. Further discussion of the Stoner I is included in the appendix.
Spin-orbit interaction is included in the form
ESO = λ~S · ~L→∑
mσ
SzLz, (3.28)
where the second form applies when only z-components of moments are treated, as is done
in current implementations of the LSDA+U method. Due to this restriction, LSDA+U
often does not produce the correct multiplet energies in the atomic limit. The visible
result in LSDA+U band structures is splittings of occupied, or unoccupied, correlated
suborbitals that can be as large as a few times J , and understanding the splittings is not
straightforward. For 4f systems these splittings[43, 44] may not be of much interest unless
one of the correlated bands approaches the Fermi level. In heavy fermion compounds, for
example, LSDA+U results are used to infer which parts of the Fermi surface has a larger
amount of f character.[45] The same effects (eigenvalue splittings) occur in 3d or 5f
systems, however, where they are expected to become more relevant but are masked by
stronger banding tendencies.
Here we consider values of λ of 0 and 0.2 eV. The magnitude of the spin-orbit interaction
is not critical to the results; it mainly serves to break degeneracies. Without the spin-
orbit interaction, the ground state for any of the functionals at a given N is degenerate
with several other states. For instance with N = 6, the AMF functional has states with
Lz = 1, Sz = 0 and L = 11, Sz = 0 with the same lowest energy.
In Fig. 3.1 are the ground states for both AMF and FLL with U = 8, J = 1 and I =
0.75. The FLL and Fl-nS (not shown) schemes both reproduce Hund’s rules exactly with
these parameters. AMF does not reproduce Hund’s rules (in fact penalizes magnetism)
until I is increased to around 1.5, which is somewhat larger than reasonable values of I. If
one expects LSDA+U to reproduce Hund’s rules, then the AMF scheme performs rather
poorly. For instance, at N = 7, Hund’s rules ask that all electrons be spin-aligned, but the
AMF ground state has only one unpaired spin due to the magnetic penalty appearing in
CHAPTER 3. LSDA+U 43
-15 -10 -5
∆EAMF
-10
-5
0
5
∆EFL
L
M = 1M = 3M = 5M = 7J = 0 values
Figure 3.2: (Color online) Shown here is ∆EFLL plotted vs. ∆EF l−S for each of the 3432configurations of N = 7 electrons, using U = 8, J = 1, I = 0.75, all in eV. The orderingof states is shown for Fl-S by counting from left to right, and for FLL by counting frombottom to top. Open squares show values for U = 7 and J = 0.
Eq. (3.13). With these parameter choices, U/(2l + 1) > I, so the AMF magnetic penalty
wins over the Stoner energy. This is likely to be the case for 3d transition metals as well,
since U3d/(2l+ 1) ∼ 1eV, but it may not be as significant since I for 3d elements is larger.
We examine the energetics in more detail in Fig. 3.2, where ∆E for the AMF and FLL
functional is plotted for every configuration forN = 7. The configurations fall into separate
lines for each spin moment M , since Edc depends only on N and M for both functionals.
For the case of J = 0, all the states with a particular M value collapse to a single energy
value (the orbital index loses any impact), this is shown with the open squares. A value
of I was chosen so that the cancellation discussed in the previous paragraph is slightly
broken.
If we examine the J = 0 case first (the large open squares in Fig. 3.2), we see that the
separation of states in FLL is much larger than AMF (9 eV versus 3 eV), with M = 7 the
lowest energy for FLL but highest for AMF. This is a direct consequence of the magnetic
CHAPTER 3. LSDA+U 44
penalty of AMF discussed previously. If I were increased above 1 eV (keeping the other
parameters fixed), then AMF would begin to favor the M = 7 state by a small amount.
Once J is turned on, the degeneracy is split, and the configurations with a particular M
spread out around the J = 0 value. The spread is especially large for the highly degenerate
M = 1 value (from -5 to 8 eV), so that even if I were larger than the typical LSDA value
(in which case, with J = 0 AMF would favor a high spin state) the large spread of M = 1
values would cause the low-spin states to be favored in AMF. This spread is entirely
coming from the EI term and is independent of the double-counting choice. Here we see
for AMF a competition between J and I: J is actually preferring a low spin configuration,
in contrast to the conventional wisdom that J increases the tendency for magnetism. We
see this same tendency occurs in FLL, as for J = 0 the separation between M = 7 and
M = 1 states is 9 eV, but with J = 1 this separation is reduced to 4 eV. Since in FLL
the Hubbard U does not penalize magnetic states the way AMF does, the presence of J is
not able to compete with I. This makes it clear why FLL is generally accepted to perform
better for systems known to have high-spin states (e.g. Eu and Gd). Conversely, FLL may
be less successful at modeling low-spin states.
As mentioned previously, it is fairly common for theoretical studies to replace U and
J with effective parameters U and J . For any double-counting term chosen, using these
effective parameters will lower the energy of the high-spin state relative to the low spin
state as compared to using U and J directly. With orbitals that are not highly localized,
such as 3d or 5f states it may be the case with FLL that the reduction of the energy sepa-
ration between high-spin and low-spin caused by using U and J would allow for significant
competition between magnetism and kinetic energy in LSDA+U.
We now have seen why and how FLL and AMF perform differently in assigning a
magnetic moment. This may be of particular interest for studies of pressure-induced
changes in magnetic moment, such as that seen in MnO[34] without changes in orbital M
occupancy. Applications of LSDA+U are more thoroughly discussed in Sec. 3.7.1.
Shown in Fig. 3.3 are scatter plots of the energies of all possible states for a given
CHAPTER 3. LSDA+U 45
0 5 10-25
-20
-15
-10
-5
0
5
10
∆E
0 5 10
N
M = 0M = 1M = 2M = 3M = 4M = 5M = 6M = 7Mean
0 5 10
(a) AMF (b) FLL (c) Fl-nS
Figure 3.3: (Color online) Scatter plot of all energies ∆E for all states in the (a) AMF(Fl-S), (b) FLL and (c) Fl-nS double-counting schemes, for U = 8, J = 1, and I = 0.75 (FLLand AMF only). Spin-orbit is neglected here. For AMF, low spin states (black circles andred squares) appear as lowest energy configurations for N near 7, but this is not the casefor FLL or Fl-nS. The dashed lines indicate the mean energy over configuration for eachN ; note that the variation with N is much less for FLL than for the other two functionals.
number of f electrons with integer occupations. SO is neglected, as it makes very minor
changes to this picture by splitting some degeneracies. The particle-hole symmetry of each
functional is apparent. In Fl-nS and FLL, the ground state energy for N = 7 is roughly
3 eV lower than the next level, which are the (degenerate) ground states for N = 6 and
8. This is almost entirely due to the term depending on M2 (either the J term in (3.8)
or the Stoner term in FLL), because M is large and at its maximum with 7 spins aligned.
In AMF low spin states can be seen at the low end of the range for configurations at
each N ; the high spin states for N=6 and 7 are disfavored by 6-7 eV. We see that the
trend where AMF favors low-spin configurations and FLL favors high-spin configurations
shown for N = 7 in Fig. 3.2 is present for all N . The large spread of values for low spin
configurations (black circles, red squares) is seen clearly for AMF as they appear in both
CHAPTER 3. LSDA+U 46
the lowest energy positions and the highest energy positions. The high-spin configurations
(large open symbols and triangles) are in the middle of each distribution for N . For e.g.
N = 5, counting from the lowest energy, M = 1 configurations are found first, followed by
M = 3 configurations then the M = 5 configurations are found (with the trend reversing
counting up to the highest energy states). In FLL, the lowest energy configurations for
N 6= 7 are still the configurations with maximum spin for a given N , and states with lower
spins are found in succession. Again using N = 5 as an example, the M = 5 configurations
are lowest in energy, and then M = 3 configurations are seen at energies lower then M = 1
states.
3.7 Discussion
In this paper we have tried to clarify the various functionals that are used in the LSDA+U
method, we have compared the functionals formally in certain limits, we have presented
the orbital potentials that arise, and we have analyzed the total energy corrections that
LSDA+U functionals apply to LSDA total energies, given a set of occupation numbers.
The Fl-nS functional which was originally introduced strongly favors spin-polarized states
as does the commonly used FLL functional. The other most commonly used functional
besides FLL, Fl-S (AMF), has characteristics that tend to suppress moment formation or
reduce the magnitude of the moment. When analyzed, this AMF functional shows positive
energy penalties to magnetism that compete with the magnetic tendencies of the LSDA
functional, and when J > 0 non-magnetic solutions become even more likely to win out.
When LSDA+U is applied to correlated insulators in the strong coupling regime, it
provides a very good picture of the system at the band structure (effective one-electron)
level. The initial successes include the 3d transition metal monoxides MnO, FeO, CoO,
and NiO, for which the LSDA description is very poor. Other early successes included the
insulating phases of the layered cuprates that become high temperature superconductors
when doped, and the unusual magnetic insulator KCuF3, which was the first case where
CHAPTER 3. LSDA+U 47
important orbital ordering was reproduced. LSDA+U is not a satisfactory theory of single
particle excitations of such systems, but nevertheless provides a realistic picture of the
underlying electronic structure.
The more interesting cases now lie between the strongly correlated limit of wide-gap
magnetic insulators and weakly correlated regime that is well described by LSDA. Some
of these are metals, some are unconventional insulators, and many lie near the metal-
insulator borderline. It is for these intermediate cases that it becomes essential, if applying
the LSDA+U approach, to understand what the method is likely to do, and especially to
understand the tendencies of the various choices of functional. This is what we have tried
to clarify in this paper. As a summary, we will provide an overview of certain results
that have appeared in the literature for systems that lie somewhere in the intermediate
correlation regime.
3.7.1 Examples of LSDA+U behavior from applications
Strongly correlated insulators.
Cuprates. The insulating phase of the cuprate class of high temperature superconductors
comprised the “killer app” that served to popularize[30, 46] the LSDA+U method, and
in the intervening years it has been applied too many times to cite. Simply put, in this
system it produces the Cu d9 ion and accompanying insulating band structure.[46, 47]
The hole resides in the dx2−y2 orbital and is strongly hybridized with the planar oxygen
pσ orbitals, as much experimental data was indicating.
MnO. Experimentally, MnO shows at room temperature a moment collapse from M =
5 to M = 1 (or less), a volume collapse, and an insulator-to-metal transition, near 100
GPa; this is the classic Mott transition. Within LSDA, the moment decreases continuously
with decreasing volume,[48] from the high spin (HS) state to a low spin (LS) state. The
insulator-to-metal transition occurs at much too low a pressure (without any other change).
A volume collapse is predicted, although the pressure is significantly overestimated (150
CHAPTER 3. LSDA+U 48
GPa).
The application of LSDA+U in its FLL flavor has been applied and analyzed in
detail,[34] and provides a different picture in several ways. The ambient pressure band
gap is improved compared to experiment. The volume collapse transition occurs around
120 GPa and is accompanied by a moment collapse from M = 5 to M = 1. The nature
of this (zero temperature) transition is insulator to insulator, while the experimental data
indicate an insulator-to-metal transition at room temperature. The zero temperature tran-
sition might indeed be insulator-to-insulator; such a phase transition would be a type that
LSDA+U should work well for. It is also possible that the static mean-field approximation
underlying LSDA+U, which favors integer occupations and hence insulating solutions, has
too strong a tendency and fails to describe this transition. This question could be settled
by studying experimentally the Mott transition at low temperature
Even more unexpected than the insulator to insulator aspect is the LSDA+U prediction
is that the low spin state has an unanticipated orbital occupation pattern,[34] being one
in which every 3d orbital remains singly occupied (as in the high spin state). but spin
in two orbitals antialign with those in the other three orbitals. This state is obtained
simply from the M = 5 HS state by flipping the spins of two of the orbitals. The resulting
density remains spherical, but the spin density exhibits an angular nodal structure leading
at the same time to a high degree of polarization of the spin-density but a low total
moment (M = 1). This solution (being the high pressure ground state in LSDA+U) can
be traced back to the interplay between symmetry lowering due to the antiferromagnetic
order (cubic lowered to rhombohedral) and the anisotropy part of the interaction Eq. 3.20).
The symmetry lowering lifts the cubic grouping (t2g and eg manifolds), thus allowing a
higher number of allowed occupation patterns.
The anisotropic part of the interaction is responsible[34] for Hund’s second rule ordering
of states, which has the tendency to increase the mutual distance of each pair of electrons.
If the all-over energetics (band broadening and kinetic effects) reduce the gain of energy
due to spin-polarization, then Hund’s first rule may become suppressed and the result is a
CHAPTER 3. LSDA+U 49
low spin state. The anisotropic interaction however, is not influenced by this suppression,
since it is a local term proportional to a parameter J . It will enforce a Hund’s second rule
like separation of the electrons under the low spin condition, which can be shown to result
exactly in the occupation pattern observed for MnO. In a sense the low spin state is an
example of Hund’s second rule without Hund’s first rule.
FeO, CoO, NiO. Together with MnO, these classic Mott (or ‘charge transfer’) insulators
have been prime applications of the LSDA+U method.[49, 50, 51, 52]. The behavior of
the open 3d shell in these compounds has not been analyzed in the detail that was done
for MnO, however.
Metals
Correlated metals involve carriers that can move, hence they invariably involve fluctua-
tions, in occupation number, in magnetic moment, in orbital occupation, etc. It cannot be
expected that a self-consistent mean field treatment such as LSDA+U can answer many
of the questions raised by their behavior. However, there is still the question of whether
LSDA+U can provide a more reasonable starting point than LSDA alone in understanding
these metals. In our opinion, this remains an open question, one for which some evidence
is available.
The Fe-Al system has provided one platform for the application of LSDA+U to moder-
ately correlated metals. The systems treated include the Fe impurity in Al (Kondo system,
experimentally), and the compounds Fe3Al, FeAl, and FeAl3. The behavior is too com-
plex to summarize here. The LSDA+U result will, generally speaking, be likely to give a
good picture of a Kondo ion when it produces an integer-valent ion with a large value of
U. Both FLL and AMF functionals have been applied,[53, 54] with substantially differing
results, leading one to question whether either is more realistic than simple LSDA. Results
are also sensitive to volume, i.e. whether using the experimental lattice constant or the
calculated equilibrium value, and the calculated equilibrium is different from LSDA and
LSDA+U. One result was that, for moderate UFe ∼ 3-4 eV, AMF strongly reduces the
CHAPTER 3. LSDA+U 50
magnetic moment, while FLL does not.[54] Another application found that the magnetism
disappeared within a certain range of intermediate values of UFe, that is, it was magnetic
around small UFe and also again at large coupling,[53] but non-magnetic between.
Moderately strongly interacting oxides.
Trying to address seriously the electronic structure of intermediate coupling oxides, which
are often near the metal-insulator transition, is a challenge that has begun to be addressed
more directly. The peculiar NaxCoO2 system, which becomes superconducting when hy-
drated (water intercalates between CoO2 layers) is one example. One set of studies showed
no appreciable difference between FLL and AMF,[55] with both predicting charge dispro-
portionation on the Co ion for x=13 and 1
2 for U ≈ 2.5-3 eV. It is likely that this compound
presents a case where the interplay between LSDA and U has effects that are not fully
understood. Also, it is unclear why there is so little difference between the FLL and AMF
functionals.
The compound Sr2CoO4 is another example. Both functionals show a collapse of the
moment[56] around U = 2.5 eV, related to the metal-half metal transition that occurs,
but the result for the moments (M(AMF) < M(FLL)) bears out the tendency of AMF
to penalize magnetic moments.. The fixed spin-moment calculation in Fig. 9 in Ref. [56]
is interesting too, showing the competition between LSDA magnetic energy and AMF
magnetic penalty. Also it shows the creation of local minima around M = integer values
that LDA+U introduces.
f electron materials.
4f systems. These metals often display the correlated electron physics of a magnetic in-
sulator at the band structure level: background conduction bands provide the metallic
nature, while the correlated states have integer occupation. The LSDA+U method seems
to be a realistic method for placing the f states closer to where they belong (away from
the Fermi level). Gd is a good example, which has been studied at ambient pressure and
CHAPTER 3. LSDA+U 51
compared to photoemission data[50] and magnetic dichroism data.[57, 58] The LSDA+U
method has also been applied up to extremely high pressure to assess where the ‘Mott
transition’ in the 4f bands is likely to occur. The LSDA+U method has also been ap-
plied to heavy fermion metals, for example Cu and U compounds,[59] PrOs2Sb12,[60] and
YbRh2Si2 [61]. In such systems the LSDA+U method may even provide a good estimate
of which itinerant states at the Fermi level are most coupled to the localized f states,
i.e. the Kondo coupling matrix elements. These 4f systems may become heavy fermion
metals (YbRh2Si2) or novel heavy fermion superconductors (YbAlB4), or they may remain
magnetic but otherwise rather uninteresting metals (Gd).
5f systems. A variety of application of the LSDA+U method to 5f systems, and
especially Pu, have been presented.[62, 63, 64, 65, 66] Given the complexity of the phase
diagram of elemental Pu, together with claims that dynamic correlation effects must be
included for any realistic description of Pu, a more critical study of Pu would be useful.
3.8 Acknowledgments
We have benefited from discussion on various aspects of this work with M. Johannes, J.
Kunes, A. K. McMahan, I. I. Mazin, and G. Sawatzky. This project was supported by
DOE through the Scientific Discovery through Advanced Computing (grant DE-FC02-
06ER25794) and by DOE grant DE-FG02-04ER46111.
3.9 Calculation of the Stoner I for 3d and 4f Shells
The Stoner parameter I is a well established quantity. For metals its value is obtained by a
second order expansion of the LSDA xc-energy around the non-magnetic solution, resulting
in a Fermi surface averaged integral of the radial wave functions with the xc-kernel.[67]
LSDA+U is usually applied to describe insulating states, where the Fermi surface vanishes.
In the context of discussing the LSDA contribution to the energy of a correlated d- or f -
shell it is more natural to consider the energy contribution from the localized shell. This
CHAPTER 3. LSDA+U 52
58 60 62 64 66 68 70Z
0.6
0.7
0.8
0.9
ISton
er [
eV] 4f
4f 3+
20 22 24 26 28 30Z
0.6
0.8
1.0
1.2
ISton
er [
eV] 3d PW92
3d X-only
Figure 3.4: Shell-Stoner integrals for the 3d and 4f atoms. For explanations see text.
CHAPTER 3. LSDA+U 53
leads to a derivation of the Stoner-I similar to the formulation of Janak but adapted to
atom-like situations.
Seo presented[68] the second order perturbation theory of the spin polarization in DFT,
which results in explicit expressions for the shell exchange parameter Inl that are applicable
to atom like situations. In this work a numerical estimate for Inl was derived indirectly
from exchange splittings and spin polarization energies taken from DFT calculations. The
idea behind this perturbation theory, the expansion the xc-energy around the spherically
averaged non-magnetic density of the shell under consideration, was also discussed in the
appendix of Kasinathan et al.[34] and leads to ∆Exc ≈ −14InlM
2 with the shell-Stoner
integral
Inl = − 1
2π
∫
K0 (r) [Rnl (r)]4 r2dr (3.29)
K0
(
~r,~r′)
=δ2Exc
δm (~r) δm (~r′)
∣
∣
∣
∣
nspher,m=0
→ K0(~r)δ(~r − ~r′). (3.30)
The last expression applies for a local approximation (viz. LSDA) to Exc. K0(~r,~r′) is
a magnetization-magnetization interaction, directly analogous to the second functional
derivative of the DFT potential energy with respect to n(~r), which is the Coulomb inter-
action e2/|~r − ~r′| plus an ‘xc interaction’ arising from Exc.
For a more detailed discussion of the parameter Inl we performed LSDA calculations
for free atoms and ions and explicitly calculated Inl from Eq. (3.29). It turns out that
∆Exc(M) given above is by far the largest M -dependent term of the energy expansion.
The spin polarization energy of isolated atoms/ions with spherical M is well described by
this estimate with an error smaller than 5 − 10%. The resulting shell-Stoner integrals Inl
have very similar values compared to the ones obtained from the theory for the metallic
situation. (Note, however, that there is a factor of 2 difference in the definition of the
CHAPTER 3. LSDA+U 54
Stoner I in some of the publications.)
For the 3d transition element series we get values Inl ranging from 0.62 eV for Sc to 0.95
eV for Zn (see Fig. 4). These values increase across the series by ≈ 0.15 – 0.20 eV, when
the exchange only LSDA is used, pointing to a reduction due to (LDA-type) correlation
effects when the full xc-kernel is used. For the 4f series the shell-Stoner integrals vary from
0.58 eV for Ce to 0.75 eV for Yb. The LDA correlation effects amount to 10% of these
values. The values obtained depend on the choice of the reference system, which serves as
zeroth order in the functional expansion. For instance for the 3+-ions of the 4f -series I4f
is increased by 6 − 20% with respect to the neutral atoms.
55
Chapter 4
Dynamical Mean Field Theory
4.1 Introduction
In recent years, Dynamical Mean Field Theory (DMFT) has gained popularity as a method
for studying correlated electron systems. There are many examples where LSDA+U fails
to adequately model materials. In DMFT, one constructs separate crystal and impurity
problems and solves them self-consistently. In the impurity problem, the electrons interact
with an effective bath, or mean field, of conduction electrons determined by the crystal
problem. The crystal problem is solved with a Green’s function approach using the self-
energy from the impurity problem. The impurity problem can be solved by any of a
number of many-body techniques designed for that purpose. There are many extensive
and detailed reviews of the DMFT method in the literature which discuss both general
DMFT and several different impurity solvers that can be used. [69, 70, 71]
4.2 Cavity Method Derivation
There are several different approaches for deriving the DMFT equations. Here I will
provide a derivation of DMFT using the cavity method, following closely the approaches
taken in Refs. [70] and [72].
CHAPTER 4. DYNAMICAL MEAN FIELD THEORY 56
The partition function from the path integral formalism of quantum mechanics is
Z =
∫
∏
iσ
dc†iσdciσe−S , (4.1)
where the integration variables are Grassman variables, and S is the action. Consider for
concreteness, the action for the Hubbard model
S =
∫ β
0dτ
∑
iσ
c†iσ(τ)
(
∂
∂τ− µ
)
ciσ(τ) −∑
ijσ
tijc†iσ(τ)cjσ(τ) +
∑
i
Uni↑(τ)ni↓(τ)
.
(4.2)
(Hereafter the τ argument to the Fermion operators will be left off for brevity, except where
necessary.) The exact nature of the interaction term is unimportant for this derivation,
except that it only contains intrasite interactions, and there are no intersite interactions.
The approach taken in the cavity method is to pick a particular site labeled with p,
and ‘remove’ it from the lattice. By integrating out the lattice degrees of freedom, we
calculate an effective action for the impurity, satisfying
1
Zeffe−Seff =
1
Z
∫
∏
i6=p,σ
dc†iσdciσe−S . (4.3)
To start with this we separate the action into the contributions Sp, containing only
site p, S(0), coming from the lattice with the cavity, and ∆S which connects the lattice to
site p. These contributions are
Sp =
∫ β
0dτ
[
c†pσ
(
∂
∂τ− µ
)
cpσ + Unp↑np↓
]
(4.4)
∆S = −∫ β
0dτ
[
∑
iσ
tipc†iσcpσ + h.c.
]
(4.5)
S(0) =
∫ β
0dτ
∑
i6=p,σ
c†iσ
(
∂
∂τ− µ
)
ciσ −∑
i,j 6=p,σ
tijc†iσcjσ +
∑
i6=p
Uni↑ni↓
. (4.6)
CHAPTER 4. DYNAMICAL MEAN FIELD THEORY 57
Then we rewrite the partition function
Z =∏
σ
∫
dc†pσdcpσ exp(−Sp)∫
∏
i6=p
dc†iσdciσ exp(−S(0) − ∆S) (4.7)
The second exponential is expanded to give
exp(−S(0) − ∆S) = exp(−S(0))(
1 − ∆S + 12!∆S
2 − . . .)
(4.8)
where if one writes ∆S =∫ β0 ∆S(τ), then powers of ∆S are evaluated with expressions
like
∆S2 =
∫ β
0
∫ β
0dτ1dτ2 T∆S(τ1)∆S(τ2) (4.9)
where T is the time-ordering operator.
Now if the expansion (4.8) is put into (4.7) this gives
Z =∏
σ
∫
dc†pσdcpσ exp(−Sp)∫
∏
i6=p
dc†iσdciσ exp(−S(0))
[
1 − ∆S +1
2!∆S2 − . . .
]
(4.10)
Z =∏
σ
∫
dc†pσdcpσ exp(−Sp)[
1 − 〈∆S〉(0) +1
2!〈∆S2〉(0) − . . .
]
(4.11)
where the integration over the lattice variables is reduced to average over the action of the
lattice with the cavity of powers of ∆S. The first-order term in this expansion is
〈∆S〉(0) = −∫ β
0dτ∑
iσ
[
tip〈c†iσ〉(0)cpσ + t†ipc†pσ 〈ciσ〉(0)
]
= 0 (4.12)
since the averaging over the lattice variables does not include averaging over site p. Simi-
larly, all odd terms in the expansion (4.11) are zero because they average odd numbers of
CHAPTER 4. DYNAMICAL MEAN FIELD THEORY 58
Fermion operators. The second-order term is
〈∆S2〉(0) =
∫ β
0
∫ β
0dτ1dτ2
∑
i,j 6=p,σ
tiptpj c†pσ(τ1) 〈Tciσ(τ1)c†i′σ(τ2)〉 cpσ(τ2)
=
∫ β
0
∫ β
0dτ1dτ2
∑
i,j 6=p,σ
tiptpj c†pσ(τ1)G
(0)ij (τ1 − τ2) cpσ(τ2). (4.13)
This, and all higher order terms in the expansion (4.7), contain unconnected Green’s
functions, meaning they can be written as products of lower-order (single particle) Green’s
functions. The linked cluster theorem allows us to compute the effective action [70, 72]
Figure 4.2: Plot of Σ(ω) for various values of the chemical potential µ, at T = 300 Kand U = 8 eV for 14 orbitals. The occupation number nat is 13 for all values of µ used,except where noted. A ‘double-counting’ value of 12U has been subtracted from Σ and µfor clarity. Note the very large response to small changes in N in the small ω region ofΣ(ω).
and the occupation of the correlated orbitals in the crystal nf are functions of µat. (nf is
of course a function of the crystal chemical potential µ, but µ is fixed by the total electron
count in the crystal.) One then iterates µat until the condition
nat(µat) − nf (µat) = 0 (4.48)
is satisfied.
This approach has a few limitations. If one downfolds the Hamiltonians to just the
correlated orbitals (or uses a model Hamiltonian without uncorrelated bands, such as the
Hubbard model), then nf = Nelec will not be a function of µat. If nf is fixed to an integer
value, then there is a range of µat for which nat = nf . Several self-enegies are plotted
in Fig. 4.2, several of which (black, blue, red, green curves) have the same nat to three
decimal places, but their self-energy is different. There are several values of µat which
CHAPTER 4. DYNAMICAL MEAN FIELD THEORY 68
satisfy the condition (4.48) but they will all have different spectra. Thus, the presence of
an extra uncorrelated band in the Hamiltonian into which charge can be transfered from
the correlated orbitals is important for this method. This issue may not be quite so bad
if one solves the impurity problem allowing each orbital to have its own site energy εi, as
in (4.27).
To examine a second possible issue with this method, we refer again to Fig. 4.2, if
Σ(ω) < 0 at low frequencies, the self-energy will behave as an attractive potential, and if
Σ(ω) > 0 then the self-energy will be repulsive. The most significant contribution to the
crystal occupation matrix will come from the low frequency regions of the self-energy, since
at high frequencies ReGcrys(iωn) = Re∑
k[iωn + µ−Hk − Σ(iωn)]−1 will be dominated
by the ω−2n behavior. Small changes in µ near integer occupations can result in a dramatic
change in the low frequency behavior of Σ(ω), going from being a strongly attractive
potential at µ = 6.95 eV, N = 13.0000 to a strongly repulsive potential at µ = 7.9 eV,
N = 13.0015. If one examines a pole-expansion of the Hubbard I self-energy (described in
Sec. 5.3) this large change in the low frequency behavior is due to a pole moving across the
Reω = 0 axis as the chemical potential is increased. If the converged value of nat is not an
integer, then the spectrum will always have a Hubbard band pinned near the Fermi energy.
For example, taking δ as a small number, at nat = 13 + δ the upper Hubbard bands will
appear just above the chemical potential, but for nat = 13 − δ the lower Hubbard bands
will be just below the chemical potential.
4.5 Hirsch-Fye Quantum Monte Carlo Impurity Solver
Quantum Monte Carlo (QMC) methods are very commonly used as DMFT impurity
solvers because they provide an essentially exact calculation of the partition function and
Green’s function for the system. QMC methods stochastically sample the Hilbert space
by generating random configurations. There are several different QMC methods used in
condensed matter physics today. The method given by Blankenbecler et al. [73] is com-
CHAPTER 4. DYNAMICAL MEAN FIELD THEORY 69
monly used on lattices, and the Hirsch-Fye algorithm [74] is commonly used as an impurity
solver for DMFT. Here I will discuss the basics of the Hirsch-Fye method, in the context
of DMFT.
Consider a single-band Hamitonian H = T + U which is separable into a kinetic part
T and an interaction part U . T must be quadratic in Fermion operators, and U must take
the form of a density-density interaction
V =∑
mm′
Umm′nmnm′ (4.49)
where the indices m,m′ may be a combined orbital-spin index. The partition function for
this system may be written
Z = Tr e−βH =L∏
l=1
e−∆τ(T +V) ≈L∏
l=1
e−∆τT e−∆τV (4.50)
with ∆τ = β/L. The breakup of the exponentials in (4.50) is the only source of systemic
error introduced into the algorithm. The first term in the error in lnZ is related to
the commutator [∆τT ,∆τV] ∝ ∆τ2, so we see that as L gets large, the error will be
proportionate to L−2, going to zero as L goes to infinity.
The Fermion operators in V can be decoupled by introducing auxillary spin fields σlmm′
with the identity
exp(
∆τUmm′(
nmnm′ − 12(nm + nm′)
))
= exp(λmm′σlmm′(nm − nm′)) (4.51)
cosh(λmm′) = exp(−∆τUmm′/2) (4.52)
These auxillary spins may take values of ±1, so they are often thought of as Ising spins. One
such spin is introduced for every coupling m,m′ at every time slice. With this decoupling
CHAPTER 4. DYNAMICAL MEAN FIELD THEORY 70
the Hamiltonian at each time slice l can be written as
Hl =∑
αβ
H lαβc
†αcβ . (4.53)
If we then define the matrix
Bl = e∆τHl
(4.54)
then the partition function can be written as the determinant of an NL×NL matrix
Z = detG−1 = det
1 0 0 B1
−B2 1 0 · · · 0
0 −B3 1 0
.... . .
0 0 0 −BL 1
. (4.55)
and with a matrix identity, (4.55) can be reduced to an N ×N matrix,
Z = det(1 +BLBL−1 · · ·B2B1). (4.56)
The QMC procedure proceeds by making flips of single Ising spins σlmm′ and accepting
or rejecting these moves based on their relative probability. When a spin is flipped at at
time slice l, this results in a change in the B matrix
Bl = e∆τHl −→ Ble
∆τ∆Hl ≡ BlΞl (4.57)
where Ξl is the identity matrix, except for an element (1+D) at the position the Hl matrix
changed when the spin flipped (henceforth denoted as γ). The probability for accepting
the spin flip is related to the ratio of the new partition function to the old one
The actual probability when a thermal bath is present (always the case in DMFT) is
Znew/(Zold + Znew) = R/(1 +R).
Once a spin flip is accepted, the Green’s function can be updated with
Gnew = (1 + (1 −G)(Ξl − 1))−1 G. (4.59)
Since Ξl−1 is a matrix with a single non-zero element, the update equation can be reduced
to
Gnewαβ,ll′ = Gαβ,ll′ −
Xαγ,ln ∆Hnγγ Gγβ,nl′
1 +Xγγ,ll′∆Hγγ(4.60)
where Greek letters inidate spacial or orbital indices, Roman letters indicating time slice
indices, and γ and n respectively indicate the site(orbital) and time slice indices where the
spin flip occurred. The matrix X is defined by X = 1 −G.
4.5.1 Analytic Continuation
The QMC algorthim calculated G(τ), the Green’s function on the imaginary time axis.
This can be easily transformed to G(iωn) where iωn are the imaginary Matsubara fre-
quencies. Generally one is interested in the real frequency spectral function A(ω) =
− 1π ImG(ω). The imaginary time Green’s function can be calculated from A(ω) with
the relation
G(τ) =
∫ ∞
−∞
e−τω
1 + e−βωA(ω) dω, (4.61)
however this relation is difficult to invert numerically due to the exponential decay in the
Kernel. In any case, G(τ) can be Fourier transformed to G(iωn), and in this case the
relation (4.61) becomes
G(iωn) =
∫ ∞
−∞
A(ω)
iωn − ωdω. (4.62)
There are several different methods for constructing A(ω) from G(τ) [75, 76, 77], but
the most successful method in recent years is the Maximum Entropy (MaxEnt) Method
[78, 79]. Here I will describe the MaxEnt method, following the nomenclature used by
CHAPTER 4. DYNAMICAL MEAN FIELD THEORY 72
Silver et al. in Ref. [78]. The basic idea behind it is that since the relation (4.61)
cannot be inverted accurately for noisy data, we construct the best A(ω) we can (based
on statistical arguments) which gives the correct G(τ).
The statistical argument is based around the Baynes theorem which states
P [X,Y ] = P [X|Y ]P [Y ] = P [Y |X]P [X] (4.63)
where P [X,Y ] is the joint probability of X and Y , and P [X|Y ] is the probability of X
given Y . In the application to QMC data, X and Y become G(τ) and A(ω). Ultimately,
the algorithm seeks to maximize the posterior function P [A(ω)|G(τ)], calculated via
P [A(ω)|G(τ)] = P [G(τ)|A(ω)] · P [A(ω)] · 1
P [G(τ)]. (4.64)
P [G(τ)] does not need to be considered here, as the probability of measuring the G(τ)
that has already been measured will not be relevant for finding the A(ω) for that G(τ).
P [G(τ)|A(ω)] is termed the likelihood function, as it indicates the probability of measuring
the measured Green’s function given that the spectral function is A(ω). To calculate the
likelihood function some assumption must be made about the error in the measured data
points for G(τ). Generally one assumes that distribution for each of the data points
G(τl) is a Gaussian distribution with standard deviation σl. (This is generally a good
assumption, however care must be taken for τ -points where the measured G(τ) is close to
zero, because G(τ) in the range (0, β) is a non-negative function, so the data measurements
at these τ -points will not likely be Gaussian distributed.) In this case, one calculates the
fitted Green’s function Gf (τl) using the relation (4.61) and the likelihood function is then
calculated as
P [G(τ)|A(ω)] ∝ exp
[
−1
2χ2
]
(4.65)
χ2 =∑
l
1
σ2l
[Gf (τl) −G(τl)] . (4.66)
CHAPTER 4. DYNAMICAL MEAN FIELD THEORY 73
The function P [A(ω)] is called the prior probability; it represents the probability of A(ω) as
a spectral function before making any QMC measurements. The prior probability should
enforce any constraints on A(ω) (for instance, P [A(ω)] should be zero for if A(ω) < 0 for
any ω). The method for calculating the prior probability is what gives this method its
name:
P [G(τ)|A(ω)] ∝ exp [αS] (4.67)
S =
∫
dω
[
A(ω) −M(ω) −A(ω) lnA(ω)
M(ω)
]
(4.68)
where S is the entropy taken from information theory. M(ω) is called the default model,
and it is used as a starting guess for A(ω). If the statistical data on G(τ) is good then the
choice of the default model should not have an impact on the resulting A(ω).
4.5.2 Orbital Susceptibility
The q-dependent orbital susceptibility can be calculated as follows.
We use a Dyson-like equation to compute the q-dependent susceptibility,
χq = χ0q + χ0
qΓχq (4.69)
where the vertex function Γ has to be calculated from the impurity solver [80] and the
bare susceptibility is given by
(
χ0q
)
ij,lm= T
∑
k
GkilG
k+qmj (4.70)
where vkl is an arbitrary potential applied to the orbitals. By choosing v in different ways,
one obtain get the magnetic susceptibility, orbital susceptibility, etc. In fact all (frequency-
independent) susceptibilities can be obtained as linear combinations of the matrix elements
of χq.
The vertex function must be calculated within the QMC algorithm. Since it is momen-
CHAPTER 4. DYNAMICAL MEAN FIELD THEORY 74
tum independent, evaluating on the impurity site is done with an equation of the same form
as (4.69), however the q dependent susceptibilities are replaced by local susceptibilities.
[81]
χloc = χ0loc + χ0
locΓχloc (4.71)
The two-particle function χloc can be calculated from one particle Green’s functions
using the Wick theorem
(χloc)ij,lm = 〈cic†jclc†m〉
= 〈cic†j〉〈clc†m〉 − 〈cic†m〉〈clc†j〉 (4.72)
where the time arguments on the Fermionic operators have been suppressed for clarity.
The bare local susceptibility is trivially calculated using the QMC local Green’s functions
(χ0loc)ij,lm = (Gloc)il(Gloc)mj .
There is no self-consistency to this procedure. Once the DMFT(QMC) calculation has
converged, one needs only to calculate the vertex function and then the full susceptibility
(4.69) can be calculated for any q-point desired.
75
Chapter 5
Charge Self-Consistent
Implementation of LDA+DMFT
5.1 Local Orbital Basis
The FPLO code uses a local orbital basis, the details of which are described elsewhere
[22]. For our purposes here it is important to realize that the basis functions are non-
orthogonal. The on-site atomic orbitals are orthogonal, but Bloch basis functions are not
orthogonalized. This necessitates the use of the overlap matrix Sk in DFT or DMFT
formalisms. Some care must be made in defining creation and annihilation operators.
Throughout this section, it will be necessary to refer to the square root or inverse of a
matrix. This will be done with the notation S1/2, and a reference to matrix elements will
be S1/2ab which means the a,b element of the matrix S1/2. A matrix element is raised to a
power will be clarified with parenthesis, e.g. (Sab)1/2.
5.1.1 Nonorthogonal Basis
First, we examine the properties of a non-orthogonal basis. We will use roman characters
to denote states in a non-orthogonal basis and greek characters to denote states in an
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 76
orthogonal basis. The overlap matrix is defined by
〈b|a〉 = Sba. (5.1)
We will define an orthogonal basis using a Lowdin transformation,
|β〉 =∑
b
S−1/2bβ |b〉 (5.2)
〈β| =∑
b
S−1/2βb 〈b|. (5.3)
It is straightforward to show that 〈β|α〉 = δβ,α. There is a one-to-one correspondance
between the states in the orthogonal basis and the states in the non-orthogonal basis, so
indices on the S matrix can be mixed without ambiguity. It is necessary, of course, that
even though S has many possible square roots, the matrix S1/2 must be chosen in a way
that is consistent throughout this paper. The matrix S−1/2 may not be an arbitrarily
chosen square root of S−1; it must be the inverse of the chosen S1/2.
The inverse transformations to (5.2) and (5.3) can be found as
|b〉 =∑
β
S1/2βb |β〉 (5.4)
〈b| =∑
β
S1/2bβ 〈β|. (5.5)
The completeness relation can be written in terms of the non-orthogonal basis states as
1 =∑
α
|α〉〈α| =∑
ab
|a〉S−1ab 〈b| (5.6)
5.1.2 Second Quantization
Now we can relate the creation and annihilation operators for the non-orthogonal basis to
the operators for the orthogonal basis. We will only consider Fermions here. Annihilation
operators will be denoted with a letter c, and creation operators with the letter d. As we
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 77
shall see shortly, for the non-orthogonal basis these are not adjoints of one another. (This
relation is, of course, still true for the orthogonal basis, ie. dβ = c†β .) The annihilation
operator can be found by expanding in operators for the orthogonal states. This expansion
gives
cb|b〉 =∑
β
Aβcβ∑
α
S1/2αb |α〉 (5.7)
=∑
β
AβS1/2βb |0〉 = |0〉 (5.8)
where |0〉 is the vacuum state. We have used the fact that cβ |α〉 = δβ,α|0〉. The solution
here is clearly that Aβ = S−1/2bβ . Then we have our annihilation operator as
cb =∑
β
S−1/2bβ cβ (5.9)
Interestingly, c†b is not an operator which creates a particle in state b. Instead, its action
on the vacuum state gives
c†b|0〉 =∑
β
S−1/2βb |β〉 (5.10)
which is not the state |b〉. In order to get a creation operator, we must find it by solving
db|0〉 =∑
β Aβdβ|0〉 = |b〉. Doing so gives the result
db =∑
β
S1/2βb dβ . (5.11)
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 78
The action of the conjugate of db is d†b|a〉 = Sba|0〉. It’s then straightforward to show that
the following anti-commutation relations hold,
cb, c†a = S−1ba (5.12)
db, d†a = Sab (5.13)
cb, da = c†b, d†a = δab (5.14)
and all anti-commutators with other combinations of c, c†, d, d† give zero.
Now we can consider how operators can be expressed in second quantization notation.
Consider a general single-particle operator H, with matrix elements Hαβ in the orthogonal
basis. The matrix elements in the nonorthogonal basis are
〈a|H|b〉 =∑
αβ
S1/2aα HαβS
1/2βb , (5.15)
which is easily obtained by inserting identity operators in the orthogonal basis. To write
the operator H in the nonorthogonal basis, first write H in the orthogonal basis, and then
use the relations (5.9) and (5.11) to expand. This gives
H =∑
ab
(S−1/2HS1/2)ab dacb
=∑
ab
dacb∑
c
S−1ac Hcb. (5.16)
5.2 LDA+DMFT
5.2.1 Double Counting
Correcting for the double counting of the Coulomb energy is done in the same way as
in LDA+U. There are two functionals that are commonly used in LDA+U, the Fully
Localized Limit (FLL) and the Around Mean Field (AMF) scheme. The LDA+U function
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 79
is generally written as
ELDA+U =1
2
∑
mm′σσ′
Uσσ′
mm′nmσnm′σ′ − EDC . (5.17)
The two double-counting terms used are
EFLLDC =1
2UN(N − 1) − 1
2J∑
σ
Nσ(Nσ − 1) (5.18)
EAMFDC =
1
2UN2 − U + 2lJ
2(2l + 1)
∑
σ
N2σ (5.19)
where l is the orbital angular momentum.
In the application of double counting to DMFT, we must apply these as orbital poten-
tials. So by differentiating the double counting energies with respect to orbital occupations
we get
V FLLσ = U
(
N − 12
)
− J(
Nσ − 12
)
(5.20)
V AMFσ =
2l
2l + 1(U − J)Nσ + UN−σ. (5.21)
For application in DMFT, one of these functionals is chosen and the potential is subtracted
from Σ(iω) for each correlated orbital.
An orbital potential that is independent of frequency could be applied in the DMFT
formalism by simply choosing Σ(ω) = const. This is essentially what LDA+U does.
The self-energy calculated from the impurity solvers differs from a constant value only at
low imaginary frequency (on the imaginary frequency axis). However, the low frequency
behavior of Gcrys plays a significant role in determining orbital occupations, so it is not
clear that the double-counting schemes used in LDA+U would have the same applicability
in LDA+DMFT. It does not appear that the details of double-counting in LDA+DMFT
have been studied in detail at all.
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 80
5.2.2 DMFT Interface with FPLO
FPLO is organized to implement the Kohn-Sham DFT formalism. This involves con-
structing the K-S Hamiltonians from the density, diagonalizing them to produce orbital
occupations, forming a new density, and repeating until the density converges. A single
step of this iteration is completed in the fplo_step function defined in the fplostep.f90
file. This function initially does some work to optimize the basis functions and set up
the Kohn Sham calculation, and then calls the function calc_KS_states to perform the
actual diagonalization of the LDA Hamiltonians. After that fplo_step computes the out-
put density and does some basic analysis of it. Mixing of the density appears to be done
at a level outside this function.
To implement DMFT, I replace the call to calc_KS_states to a call to my own function
dmft_driver. This is done with appropriate conditional statements to the application of
DMFT can be turned on/off with a switch in the input file. The function dmft_driver
then performs the entire DMFT calculation, and fills the appropriate output variables for
computing the density.
5.2.3 Calculation of Local Orbital Occupation Values
Details on the FPLO basis set can be found in Ref. [22] and will not be presented here,
except where necessary. It is worthwhile to mention that the on-site orbitals on an atom
are orthogonal, that is∑
k Sk = I, the identity matrix. Provided the crystal under study
has inversion symmetry, then for all local quantities the block corresponding to a group
of orbitals on a single atom will be diagonal (there may be off-diagonals if there are
multiple atoms in the unit cell). Off-diagonal elements in the bath Green’s function (due
to multiple atoms or numerical precision) would be neglected with an impurity solver other
than Hubbard I. Currently only Hubbard I is implemented.
The crystal Green’s function for a non-orthogonal basis is defined on the imaginary
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 81
frequency axis by (4.20)
Gk(iωn) = [(iωn + µ)Sk −Hk − Σ(iωn)]−1 (5.22)
Gnetcrys(iωn) =
∑
k
Gk(iωn) (5.23)
where iωn are the Matsubara frequencies, µ is the chemical potential, Hk is the LDA
Hamiltonian and Σ(iω) is the self-energy. The Hk and Sk matrices are of dimension
Norb×Norb, and they are Hermitian but not sparse. The Σ(iωn) matrix is also of dimension
Norb×Norb, but it has non-zero values only on the subset of the orbitals which are treated
as correlated and (usually) is diagonal. Σ(iωn) is a complex symmetric matrix and is not
Hermitian. It should also be noted here that the summation∑
k indicates an average over
k.
In the FPLO basis, (5.23) actually corresponds the on-site Green’s function (called ‘net’
in FPLO terminology), which when summed over iωn gives only the on-site contributions
to the occupation. To calculate the total occupation, one computes the gross Green’s
function
Ggrosscrys (iωn) =
∑
k
Gk(iωn)Sk (5.24)
The CPU time for evaluating this is linear in the number of Matsubara frequencies and
k-points, so very little can be done to reduce the cost in this respect. The matrix inversion
is somewhat costly, scaling as O(N3orb). This cost can be mitigated somewhat by using a
geometric series expansion for high frequencies,
Gnetcrys(iωn → ∞) =
1
iωn
∑
k
[
S−1k − 1
iωnS−1
k Lk
]
(5.25)
with Lk = µ − S−1kHk − S−1
kΣ(∞). The product S−1
kLk is independent of frequency, so
in this way the matrix inverse at each frequency can at large frequencies be replaced by
matrix addition operations which are O(N2orb).
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 82
The density matrix is calculated from Gcrys by the following formula,
ρgross = T∑
n
Ggrosscrys (iωn)e
iωn0+(5.26)
= T∑
n
Ggrosscrys (iωn) + 1
2INorb(5.27)
where INorbis the Norb ×Norb identity matrix. The density matrix can be calculated for
both net and gross occupations, which is a useful when Σ(iωn) = 0 as check of the code
against the LDA output, but the gross occupation is the physical number of electrons and
µ must be adjusted so that Tr ρgross = Nelec. Henceforth, the density matrix for gross
occupations will simply be referred to as ρ, and Ggrosscrys will be referred to as Gcrys.
The extra term in (5.27) comes from the fact that we are using the Fourier transform
Gcrys(iωn) to evaluate limτ→0+ Gcrys(τ) . Gcrys(τ) is anti-periodic with a discontinuity at
τ = 0. The jump in Gcrys at τ = 0 is 1, coming from the Fermion commutation relation. It
is well known that a Fourier series of a discontinuous function will converge to the average
of the two values on either side of the discontinuity, that is
∑
n
Gcrys(iωn) = 12
[
Gcrys(τ = 0+) +Gcrys(τ = 0−)]
(5.28)
so the extra 1/2 must be added to give the correct value.
The sum over Matsubara frequencies formally must be carried out from −∞ to ∞. Of
course, this cannot be done in practice on a computer, so we use the following optimization.
First, we can trim the sum from 0 to ∞ by noting that Gcrys(−iωn) = G∗crys(iωn), so that
ρ = 2T∞∑
n=0
ReGcrys(iωn) + 12INorb
. (5.29)
At large ωn, ReGcrys ∝ ω−2n so by fitting the line ln |ReGcrys| = A + B lnωn to find
the A coefficient the high frequency terms can be summed to infinity analytically. Also
the comparison of the B coefficient to −2 gives a measure of the quality of the assumption
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 83
of frequency being “high enough” to use this kind of fit. If the expansion in (5.25) is used,
then the intercept A can be computed directly as∑
k S−1k Lk.
The sum of this tail function is
∞∑
n=−∞
1
ω2n
= β2∞∑
n=−∞
1
[π(2n + 1)]2=
1
4β2. (5.30)
Then we compute the density as
ρ =1
2INorb
+ 2T
Nω∑
n=0
ReGcrys(iωn) +1
4AT − 2AT
Nω∑
n=0
1
ω2n
. (5.31)
Here, we have taken the analytic sum from 0 to ∞ of ω−2n , added the sum of the Green’s
function over the Matsubara frequency range used in the calculation and subtracted the
sum of ω−2n over those same Matsubara frequencies. The 1
4AT term is added to each
element in the matrix, not just along the diagonal. We note that this correction is rather
small, since summing only 128 (positive frequency) terms in (5.30) gives 0.249604β2 , a
value which is converged to 3 decimal places. The choice of the actual number of Matsubara
frequencies used in the calculation is determined by the temperature; a value should be
chosen so thatGcrys(iωn) at the highest calculated frequencies does not significantly deviate
from ω−2n behavior. At present, having the highest frequency at around 6 Hartrees seems
to be adequate. As a rule of thumb, one should choose Nω so that its product with T
measured in K is around 106.
To get meaningful results, one must adjust the chemical potential µ to satisfy Nelec ≡
Tr(ρ) = Nproton. Since the calculation of ρ requires recomputation of Gcrys at each fre-
quency, this is fairly expensive. So we must perform the search for µ in as intelligent a
manner as possible to minimize the number of iterations performed here.
First, the correct number of electrons must be bracketed by lower and upper values of
the chemical potential, µL and µH . There is in general no clear way to do this, however
we are saved in this case by the fact that Ne(µ) is a monotonically increasing function.
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 84
Once the correct electron number is bracketed, then µ must be iterated to get the
correct value. There are three kinds of updates used.
1. Bisection
µBnext =1
2(µL + µH) (5.32)
2. Linear interpolation
m = (NH −NL)/(µH − µL) (5.33)
µLnext = µL + (Np −NL)/m (5.34)
3. Average
µAnext =1
2
(
µBnext + µLnext
)
(5.35)
4. Quadratic Interpolation
Details in Sec. 5.4
Linear or quadratic interpolation are generally the best, however if there is a gap in the
electronic structure between the two brackets, then the both methods can spectacularly
fail to find the correct chemical potential. If one of the brackets is in the gap, then
interpolation schemes will continually pick points inside the gap and converge very slowly.
In this case doing an averaging update for a single step does very well to move the lower
bracket close to the target value, so that the linear interpolation will converge must faster.
5.2.4 Computation of Crystal Green’s Function
The computation of the Gcrys(iω) is performed simply as a straight sum over k-points.
This is done without regards to any crystal symmetries, so to get meaningful results,
one must generate k-points in the entire Brillouin zone, not just the irreducible zone. If
this calculation could be done in the irreducible zone, perhaps by using the symmetry
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 85
operations to symmeterize the resultant Gcrys then this could speed up the calculation
tremendously.
The difficulty here is in understanding how exactly to apply the symmetries to an
orbital-basis matrix. Symmetries are typically written as 3×3 real space matrices, perhaps
with a translation vector for non-symmorphic symmetries. How to apply this transforma-
tion in a general way to a matrix written in an orbital basis is non-trivial.
Convergency with respect to k-point sampling could likely be improved by implement-
ing a tetrahedron algorithm for the summation over k. This is not done at the moment,
but I don’t forsee any difficulty above and beyond the implementation of the algorithm
itself.
5.2.5 Mixing
Briefly here I will describe Anderson mixing as used in my DMFT implementation. Ander-
son mixing seeks to minimize the difference between averaged input and output vectors,
defined at iteration n by
|I(n)〉 = (1 − γ)|I(n−1)〉 + γ|I(n)〉
|F (n)〉 = (1 − γ)|F (n−1)〉 + γ|F (n)〉. (5.36)
Then the function ∆(n) = 〈I(n) − F (n)|I(n) − F (n)〉 is minimized with respect to γ. This
leads to the following definitions,
|M (n)〉 = |F (n)〉 − |I(n)〉
|N (n)〉 = |M (n)〉 − |M (n−1)〉
γ =−Re 〈M (n−1)|N (n)〉
〈N (n)|N (n)〉 (5.37)
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 86
The input for the next iteration is then taken as
|I(n+1)〉 = (1 − α)|I(n)〉 + α|F (n)〉 (5.38)
where α is a mixing parameter chosen empirically, and is analogous to the mixing parameter
used in simple mixing. The algorithm could be expanded to mix in iterations previous to
the (n − 1)’th iteration by introducing extra mixing variables analogous to γ. However,
the mixing as it stands is satisfactory, and it is not clear whether the convergence would
be significantly improved by doing this enough to justify the extra coding effort and the
increase in code complexity and required memory storage.
The evaluation of the inner products above can be done in any manner which is done
consistently. Some care should be taken in choosing how to perform the inner product
with self-energy matrices. Since the self-energy tends to a constant value proportional to
N at high frequencies, a method should be chosen which will not place too much weight
on this constant value. The form of the inner product used is
〈A(iωn)|B(iωn)〉 =∑
i,j,n
A∗ij(iωn)Bij(iωn)
ω20
ω20 + ω2
n
(5.39)
where ω0 is chosen at some reasonable value where Σ has not yet reached its high frequency
value.
5.2.6 Charge self-consistency
Most DMFT implementations do not complete the charge self-consistent loop. In some
cases this is due to using an impurity solver which is very computationally expensive,
such as QMC / Hirsch-Fye. The Hubbard I impurity solver is cheap enough that charge
self-consistency can be implemented. The approach described in this section could be used
for a general self-energy. However, the pole expansion described in section 5.3 provides a
much more straightforward means of implementing charge self-consistency with Hubbard
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 87
I in FPLO.
The density in FPLO is divided up into contributions from core-core terms, valence-
valence terms, and core-valence terms.
nvv =∑
k
wkV DkD†kV
† (5.40)
nvc = −∑
k
wkV DkD†kSvcC
† (5.41)
ncv = −∑
k
wkCScvDkD†kV
† (5.42)
ncc =∑
k
wkCScvDkD†kSvcC
†. (5.43)
Dk is stored in the code as the variable kohnsham%p_ccvk. In the Kohn-Sham for-
malism D is the eigenvectors of the Hamiltonian, obtained from the secular equation
HkDk = εkSkDk. The weights wk are calculated from the tetrahedron integration scheme
to properly perform the sum over k in the irreducible Brillouin zone. They are stored in
the code in the variable kohnsham%p_bweitk.
In order to trick the density computation routines defined in density.f90 into com-
puting the correct density from the valence occupation matrix calculated in DMFT, we
need to compute values of wk and Dk such that∑
kwkDkD†k = ρ. The most trivial way to
do this is to assume the elements of Dk are entirely real, and then compute the square-root
of the density matrix, giving equal weights to all k-points. In this method, D2k = ρ and
wk = 1/Nk.
This method is problematic if ρ should have negative eigenvalues, which can occur in
a non-orthogonal basis. In this case there is no way to compute a matrix satisfying both
D2 = ρ and D = D† because D must have complex eigenvalues. One would normally
expect the density matrix to be positive semi-definite, however because the basis functions
are non-orthogonal, it is possible for orbital occupations in this basis to be below 0 or
above 1, albeit by only a small amount. One possible resolution for this is to diagonalize
ρ and fix any negative eigenvalues to zero. As it is most probably in a well converged
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 88
calculation that any negative eigenvalues will be very small, it is unlikely that this fix have
a significant impact on the result, while allowing one to compute D from the squareroot
of ρ.
There is no general solution to DD† = ρ unless ρ is positive semi-definite. For a
concrete example, let ρ be a 1x1 matrix with the single element -1. It is impossible to find
a complex number z such that zz∗ = −1. Conversely, for any matrix A, the matrix AA†
is always positive semidefinite.
There is an extra degree of freedom which has been ignored above, namely the weights
wk. Through a suitable choice of wk and Dk matrices, it should be possible to find a
solution to∑
kwkDkD†k
= ρ for all possible ρ. However, a clear perscription for doing
this is not obvious. Unless the number of k-points is very small, then it is likely that
there are many more degrees of freedom here than necessary. Since any solution should
be acceptable, the key to using this approach is to find a solution which is numerically
robust.
5.3 Pole Expansion of Self-Energy
Using a numerical representation of the self-energy is very computationally expensive as it
requires a matrix inversion for every k-point, every Matsubara frequency at every chemical
potential checked when searching for the particle number. A much faster algorithm can
be developed if the self energy can be represented as a sum of poles, ie.
Σab(ω) = Σ∞ +
Npole∑
i=1
W †aiWib
ω − Pi(5.44)
where Wai applies an appropriate weighting factor. For this section, we assume Σ∞ in-
cludes any necessary double-counting correction.
The problem can be linearized with respect to frequency by using the matrix inversion
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 89
identity
(ω + µ)Sk −Hk − Σ∞ W †
W ω − P
−1
=
[(ω + µ)Sk −Hk − Σ(ω)]−1 · · ·...
. . .
(5.45)
where the matrix P is a diagonal matrix with the poles Pi along the diagonal. Because
the left hand side in (5.45) is linear in frequency, rather than invert the matrix at each
frequency we can solve an eigenvector problem to invert the matrix. If we define
Lk =
Hk + Σ∞ W †
W P
(5.46)
Ok =
Sk 0
0 I
(5.47)
then solve the generalized eigenvalue problem Lk|ψkm〉 = εkmOk|ψkm〉, then a transfor-
mation to the basis of eigenvectors can be done with the unitary matrix ckij = (ψki)j .
Then the Green’s function for the net population associated with the matrix Lk can be
calculated via
GLij(ω) =∑
km
ck∗imckmj
ω + µ− εkm. (5.48)
Actually, there is no real need to calculate the Green’s function directly. With the pole
expansion, from Hubbard I the occupation matrix can be evaluated analytically, since
∑
n [iωn + µ− ε]−1 = f(ε − µ), the Fermi function. This way, the net occupation of the
i-th orbital can be calculated as
Ni,net =∑
km
ck∗imckmif(εkm − µ) (5.49)
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 90
and the gross occupation is
Ni,gross =∑
kmj
ck∗im(Ok)ijckmjf(εkm − µ) (5.50)
This formalism increases the size of the “effective Hamiltonian” Lk and the Green’s
function by the number of poles in the expansion. The elements in the N × N LDA
Hamiltonian and its eigenvectors correspond to specific orbitals. The first N elements in
the Lk and GL matrices correspond to these same orbitals, however there are Npole extra
“orbitals” which correspond the poles that were added. These orbitals don’t have any
physical significance, so the section of the Green’s function corresponding to these orbitals
can be ignored. The increased size of Lk also means there will be extra bands, (each
eigenvalue/vector pair corresponds to a band) but since the elements that correspond to
the pole states don’t have physical significance, each band will not contribute an integral
number of electrons to the orbitals.
The implementation of this approach in FPLO provides a more straightforward means
of implementing charge self-consistency than the numerical approach. The diagonaliza-
tion of of Lk provides the right eigenvectors, and the code is already situated to handle
eigenvectors for a differing number of bands than the number of valence orbitals.
5.4 Parabolic Guess for Chemical Potential
The purpose of this section is to briefly describe how to choose an input value for an
algorithm based on a parabolic fit to three points given a target output value. The three
points will be labeled with ordered (xi, yi) pairs, so that xi < x(i+ 1).
First we define a coordinate system where a point is chosen to be the origin. For
concreteness, this will be the i = 0 point, but there is no loss of generality in making this
CHAPTER 5. CHARGE SELF-CONSISTENT LDA+DMFT 91
choice. This local coordinate system is given by
si = xi − x0 (5.51)
ti = yi − y0. (5.52)
We fit an equation of the form t = as2 + b′s with the following matrix multiplication
a
b′
=
1
s21s2 − s1s22
s2 −s1−s22 s21
t1
t2
(5.53)
then an equation of the form y = ax2 + bx+ c can be formed by calculating
b = b′ − 2ax0 (5.54)
c = ax20 − b′x0 + y0 (5.55)
and a guess value for x given a target value yt, can be calculated via
x =−b2a
± 1
2a
√
b2 − 4a(c− yt). (5.56)
One of the two possible guess values is chosen based on whether or not it’s between the
two points which bound the target value.
92
Part II
Applications
93
Chapter 6
Wannier Functions in LiNbO2
The work described in this chapter was done in collaboration with Warren E. Pickett and
published in reference [26].
6.1 Introduction
The quasi-two-dimensional (2D) compound LiNbO2 has become of interest in condensed
matter studies in no small part due to the existence of superconductivity[82] in the de-
lithiated phase LixNbO2. This system consists of a triangular lattice of transition metal
(niobium) ions separated by layers of O ions from ‘blocking’ (Li) layers, thus possessing
structural similarities to high temperature superconducting cuprates. Important differ-
ences from the cuprates include the presence of a 4d rather than 3d ion, and of course
the triangular rather than square lattice. The trigonal prismatic coordination of Nb with
six O atoms causes the valence band to be composed solely of Nb dz2 states. Other d
states are well separated by a 2 eV gap, leaving only a single band per formula unit re-
quired for analysis. The comparative structural simplicity of LiNbO2 makes it attractive
for theoretical studies, especially as it requires modest computational time to characterize
its properties.
This alkaliniobate is also of interest due to the increasing attention garnered by nio-
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 94
bates in recent years.[83] The three-dimensional LiNbO3 is an important and well stud-
ied ferroelectric material.[84] Superconductivity at Tc ∼ 20 K has been reported by a
few groups[85, 86, 87] in the barium niobate system. The phase has not been positively
identified but it has been suggested to be a nonstoichiometric BaNbOx cubic perovskite.
Superconductivity in the 5-6 K range has been reported in the Li-intercalated layered
perovskite system LixAB2Nan−3NbnO3n+1 [A = K, Rb, Cs; B = Ca, Sr, Ba; n = 3 and
4].[88, 89] This system has clear structural motifs in common with the peculiar Nb12O29
system, where the ‘single O2− deficiency’ (with respect to insulating Nb2O5 = Nb12O30)
releases two carriers into the cell; one localizes and is magnetic while the other is itinerant
and results in conducting behavior.[90, 91]
While there is good agreement as to structure type and lattice parameters of the stoi-
chiometric phase LiNbO2, there is some disagreement in the literature as to its electronic
and magnetic properties. Originally it was reported as being paramagnetic with strong
field dependence,[92] however calculations predict it to be a diamagnetic semiconductor.
[93] Other experimental work suggests that it is diamagnetic, and strongly field-dependent
paramagnetic properties are the result of impurities.[94] The values of the band gap are
also in mild dispute. Extended Huckel calculations with a single isolated NbO2 layer give
a bandgap value of 1.4 eV, [93] and full potential linear muffin tin orbital calculations with
the full crystal structure of LiNbO2 give 1.5 eV.[95] Measured optical reflectance shows an
onset at ∼ 2 eV, which is suggested to be a direct bandgap responsible for the burgundy-
red color of LiNbO2.[94] Band structure calculations given here and by Novikov et al.[95]
are in agreement that LiNbO2 is a direct gap semiconductor.
Geselbracht et al.[82] first reported that the de-lithiated phases LixNbO2 with x = 0.45
and x = 0.50 support superconductivity at temperatures below 5.5 K and 5 K, respectively.
Moshopoulou, Bordet and Capponi[96] later reported samples with x ranging from 0.69
to 0.79 showing the onset of the Meissner effect at 5.5 K, but samples with x = 0.84 and
above do not exhibit any superconducting transition down to 2 K. Superconductivity has
also been reported in hydrogen-inserted HxLiNbO2, in which both samples reported had
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 95
the same Tc = 5.5 K for x = 0.3 and x = 0.5. [97]
Removal of Li has the effect of adding holes to the conduction band made up of Nb dz2
states. As mentioned above, there does not appear to be any notable dependence[82, 96] of
Tc on the concentration of Li in the range 0.45 < x < 0.8. Such concentration independence
has also been observed in the 2D electron-doped system[98, 99] LixZrNCl. For a phonon-
paired superconductor with cylindrical Fermi surfaces this behavior is understood in terms
of phase space restrictions on phonon scattering processes[100] in 2D. The independence
of Tc on carrier concentration in this system is quite surprising however, since the density
of states (DOS) plots presented here and by Novikov et al.[95] vary strongly with energy,
whereas parabolic bands lead to a constant DOS in 2D. There has also been some work
done investigating the ion mobility of LiNbO2 to assess its potential usefulness as a battery
material.[94, 101]
In this work we study the electronic structure and bonding of stoichiometric LiNbO2 us-
ing density functional theory in the local density approximation (LDA). Density functional
perturbation theory is used to obtain zone center phonons and Born effective charges. The
structure contains a single parameter z that specifies the position of the O layers relative
to the Li and Nb layers, and which seems to be of particular importance in understanding
the electronic bonding and electron-phonon coupling in this compound. We investigate
the corresponding Raman active vibrational mode (beating oscillation of the O planes)
and the effect this has on the band structure. We obtain a tight-binding (TB) model that
involves many neighbors and helps one to understand how this variation affects interac-
tions between the Nb atoms. Many aspects of the electronic structure, such as chemical
bonding and effective charges, can be seen to be interrelated by examining the Wannier
functions of the valence bands.
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 96
6.2 Structure
Structural information of LiNbO2 has been reported by several experimental groups, lead-
ing to the assignment of space group P63/mmc (No. 194), with Li occupying sites 2a
(0,0,0) with symmetry 3m, Nb occupying sites 2d (23 ,
13 ,
14) with symmetry 6m2, and O
occupying sites 4f (13 ,
23 , z) with symmetry 3m. The O sites include an internal parameter
z specifying their height relative to the Li layer along the c-axis. The Li sites are centers
of inversion; the Nb sites have z-reflection symmetry in addition to the 3m triangular
symmetry within the layer but do not lie at centers of inversion.
There is good experimental agreement on lattice parameters, a = 2.90A, c = 10.46A.
[96, 94, 92] The distance a between Nb atoms is quite close to the nearest neighbor distance
of 2.86 A in elemental bcc Nb, suggesting that direct Nb-Nb coupling should be kept in
mind. However, Nb3+ 4d orbitals will be smaller than in neutral Nb. In later sections,
we will discuss details of the Nb-O interactions, and show that the electronic properties of
LiNbO2 are very sensitive to changes in the Nb-O distance. The vertical distance between
Nb and O planes can be calculated by (14 −z)c, and the distance between Nb and O atoms
is given by√
29a
2 + (14 − z)2c2.
Some disagreement exists about the value of z. A larger value of z indicates the O layers
are closer to the Nb layers. Meyer and Hoppe[92] found z = 0.1263, Moshopoulou, Bordet
and Capponi[96] report z = 0.128, Geselbracht, Stacy, Garcia, Slibernagel and Kwei[94]
report z = 0.12478, Tyutyunnik, Zubkov, Pereliaev and Kar’kin[102] report z = 0.1293.
These variations amount to a difference in O layer position of about 0.05A.
Optimization of unit cell geometry within LDA leads to a = 2.847A, c = 10.193
A, z = 0.1211. This differs from the measured value by ∆a ≈ −1.8%, ∆c ≈ −2.5%,
for a total volume discrepancy of ∆V/V ≈ −6%. The calculated overbinding of the
unit cell is somewhat larger than is typical in LDA calculations, perhaps due to the low
(practically zero) valence electron density in the Li layer. In this regard, we note that our
Nb pseudopotential does not include the 4p states in the valence bands, which may have
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 97
Γ M K Γ A L H A-5
-4
-3
-2
-1
0
1
2
3
Ene
rgy
(eV
)
Figure 6.1: (Color online) Band structure for LiNbO2 with O height z = 0.1263 typical ofthe slightly scattered experimental results. Two Nb bands (one per Nb layer) lie withina 5.5 eV gap between the O 2p bands below and the remaining Nb 4d bands above. Thetight binding fit of Table 6.1 is shown with a dashed line.
some effect on structural properties. Optimization with respect to z, holding the lattice
parameters a and c fixed at experimental values gives z = 0.125, which is very close to the
value of Geselbracht et al.[94]
6.3 Electronic Structure Calculations
The present results have been obtained through the use of the abinit code.[103] The lattice
constants and the parameter z given by Meyer and Hoppe[92] were used in the calculations
in this section. Pseudopotentials used were LDA, Troullier-Martins type,[18] generated
using the Perdew-Wang LDA exchange-correlation functional.[13] A self-consistent density
was generated with 36 irreducible k-points, which was then used to calculate energies at
116 k-points along high symmetry directions to plot the band structure. The kinetic energy
cutoff for the planewave basis was set to 40 Hartrees.
Figure 6.1 of the important region of the band structure shows the top of the O 2p
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 98
Figure 6.2: (Color online) Density of states for Nb and O atoms. For LiNbO2, states areoccupied up to energy E = 0. The dominant states in the valence band for Nb are 4d-orbitals, and for O they are 2p-orbitals. Dashed vertical lines are included to indicate theFermi energy in the rigid band picture for LixNbO2 for concentrations x = 0 (half-filledband), x = 0.4, x = 0.8 from left to right. The last two values of x roughly bracket theconcentrations where superconductivity is observed.
bands, a 2 eV gap to two Nb 4d bands which are the highest occupied bands, and another
2 eV gap to the conduction bands. The band gap, direct at the Γ point, is 1.9 eV,
somewhat higher than the 1.5 eV reported by Novikov et al. (also at the Γ point) using
the full potential linear muffin-tin orbital method (FLMTO) with the same structural
parameters.[95] LDA calculations often underestimate band gaps, so we conclude that it
is most likely the experimental band gap is somewhat larger than 2 eV. Measured optical
absorption suggests that the band gap should be around 2 eV,[94] which is consistent with
the reported dark reddish color.[94, 92]
Projected density of states (DOS) calculations, presented in Fig. 6.2 were done using
the tetrahedron method projecting to atomic spheres with radii of R(Li) = 1.7 a.u., R(Nb)
= R(O) = 2.0 a.u. These calculations confirm that the uppermost valence bands are largely
Nb 4d character, with roughly 15% contribution from O 2p states, and virtually none from
Li states. These projected DOSs are consistent with those calculated by Novikov et al.[95]
The trigonal crystal field splitting of the Nb 3d states results in the dz2 orbital energy
centered 4 eV below the other Nb 4d states (two eg doublets, based on xz, yz and
x2 − y2, xy). The unit cell contains two symmetry-related Nb layers which result in the
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 99
presence of two Nb dz2 bands in the band plot. The splitting of the bands along Γ-M-K-Γ
reflects the strength of interaction between layers, as discussed below.
As a result of the large trigonal crystal field splitting, the Nb dz2 state forms a single
band, triangular lattice system of the type that is attracting renewed interest. The cal-
culated DOS in Fig. 6.2 has only a vague resemblance to the nearest neighbor TB model
(see for example Honercamp[104]) making it evident that longer range hopping must be
important. We address this question in detail in section 6.6.
This general electronic structure, when hole-doped, is rather remarkable. It is a bona
fide representation of a single band system on a triangular lattice, moreover the band is
well isolated from other bands. Although single band models are widely used to investigate
concepts and to model many-body effects, there are very few actual realizations in real
solids. The value of the effective Coulomb on-site repulsion U for Nb3+ is not established,
but with W < 1.75 eV it will not require a very large value to lead to correlated behavior.
The hole-doped system LixNbO2 provides a unique platform for the detailed comparison
of single band models with experimental data, which for thermodynamic properties and
for spectroscopic data below 2 eV should involve only the single active band.
6.4 Zone Center Wave functions
We can examine Kohn-Sham wavefunctions to gain insight into the electronic structure.
Shown in Fig. 6.3 are isosurface plots of the Γ wavefunctions for the lower and upper
valence bands. Both plots reveal the dz2-character of the bands on the Nb atoms, as well
as 2pz bonding contributions from the O atoms. For the upper band, the p-lobes have more
weight away from the sandwiched Nb layer, indicating the weight of bonding with O is
somewhat decreased. The change in sign between the two Nb layers in Fig. 6.3b indicates
the presence of a nodal plane in the Li layer arising from the wavefunction’s antibonding
character (with respect to interlayer coupling).
The presence of an interesting triangular shape of the wavefunction around the waist
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 100
Figure 6.3: (Color online) Isosurface plot of the Γ-pt wavefunctions for LiNbO2 for (a)the lower of the pair of valence bands, and (b) the upper valence band, made withXCrysDen.[105] Both wavefunctions are generated with the same isosurface value. Nbatoms are turquoise (grey), O atoms are dark red (dark gray), Li are light gray. Yellowand blue (light and dark) indicate opposite signs of the wavefunction. The z2 symmetryaround the Nb atom is apparent as well as the contribution to the bonding made by theO atoms. Trigonal prisms with O atoms at the corners and Nb atoms in the center havebeen outlined as a guide to the eye.
of the Nb atoms is more unexpected, being noticeably larger for the bonding function.
The corners of these triangles do not point toward nearest neighbors, instead they point
toward interstitial holes in the Nb layer which are not directly sandwiched between O
atoms. This shape reflects either three-center Nb-Nb bonding, which would result in an
increased density in this region, or antibonding character of direct Nb-Nb interactions that
would decrease the density along the directions connecting Nb atoms. The character and
origin of this triangular ‘waist’ around Nb atoms are clarified in the following section.
6.5 Wannier Function
A single isolated band also provides an unusually clean system for performing the trans-
formation from reciprocal space (bands) to real space (bonds) to look at character and
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 101
Figure 6.4: (Color online) Isosurface plot of the Wannier function for LiNbO2, made withXCrysDen.[105] Niobium atoms are turquoise (gray), O atoms are dark red (dark gray),Li not pictured. Yellow and blue indicate opposite signs of the WF. The z2 symmetryaround the central atom is apparent, as well as the xy/x2 − y2 character around nearestneighbors, which maintains the orthogonality of WFs centered on different atoms.
strength of interatomic interactions from a local viewpoint. The straightforward way of
doing this is to generate Wannier functions (WFs) defined by
|Rm〉 = N−1/2k
∑
k,n
Ukmne
−ik·R|kn〉 (6.1)
where Uk can be any unitary matrix, and Nk is the number of k-points used in the
summation. The orthonormal Bloch states are denoted by |kn〉. If there were only one
NbO2 layer per cell, hence only one band, U(k) would be simply a complex number of
unit modulus. Here it is a 2× 2 matrix in the band space, which gives rise to very modest
extra complexity. These WFs display orthonormality 〈Rm|R′m′〉 = δR,R′δm,m′ . Once a
choice for Uk is made, this approach generates a basis set of identical WFs for each point
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 102
Figure 6.5: (Color online) Isosurface plot of the difference |Wz=0.1359(r)| − |Wz=0.1263(r)|,looking along the [010] cartesian direction. Yellow (light gray) is the positive isosurfaceand blue (dark gray) is negative, with Nb atoms in turquoise (gray) and O atoms in darkred (dark gray). As the O atoms are moved closer to the Nb layer, there is an increase indensity above and below the Nb atom; in addition the contributions from the O-p orbitalsappear to tilt more towards the central Nb atom.
in the Bravais lattice of the system being studied.
We have chosen the unitary matrix U such that U = M(M †M)−1/2 with Mmn =
〈kn|Sm〉 where |Sm〉 are a set of atom-centered functions with the desired dz2symmetry.
This technique has been used previously in the study of magnetic ordering in cuprates.[25]
Choosing |Sm〉 as hydrogenic 4dz2 orbitals centered on the Nb atoms for the two entangled
valence bands gives a pair of identical WFs for each Nb bilayer in the unit cell, based on
the Nb d2z symmetry that was also observed in the Γ-point wavefunction. An 8 × 8 × 4
mesh totaling 256 k-points was used to generate the WF shown in Fig. 6.4. Increasing the
size of the mesh does not change the visual appearance of the WF.
The WF naturally displays the prominent lobe along the ±z axis that is characteristic
of the 3z2 − r2 function. It also contains pσ contributions from the nearest O ions, as well
as smaller but still clear 2p amplitude on the second neighbor O ions that is approximately
pπ. Here pσ denotes the combination of px, py, pz functions oriented toward the reference
Nb atom, while pπ is perpendicular. The pσ lobe nearest the Nb ion has the same sign as
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 103
the 3z2 − r2 lobes along the z axis.
The in-plane part of the WF reveals more about the bonding in LiNbO2. The ring-
structure of the 3z2 − r2 function in the x − y plane connects with combinations of xy
and x2 − y2 functions on the three neighboring Nb ions to form a substantial in-plane
windmill structure with three paddles, which are oriented between the z-projections of the
neighboring O ions. A “3z2 − r2” symmetry WF has the full D3h symmetry of the Nb
site, which allows a contribution of an appropriate combination of x2 − y2 and xy atomic
orbitals, and the windmill structure reflects the bonding of such functions on neighboring
Nb ions. Thus the 4d character is not entirely dz2 , which clarifies the threefold symmetric
shape around the Nb waist in the k = 0 wavefunction discussed in the previous section.
In addition, there is a ‘hotdog’ structure of opposite sign connecting pairs of nearest-
neighbor Nb ions. Of the two types of Nb-Nb pairs, these hotdogs lie nearest to the first
neighbor O ions. Small nearest-Nb “backbonds” in the x − y plane of both the windmill
arms and the hotdogs are also evident. As we note below, the WF contains information
about the Nb-Nb hopping parameters. The more distant parts of the WF function are a
direct reflection of the existence of such longer range interaction.
This Nb 4dz2 WF is similar to that for Ta 5dz2 in isostructural TaSe2.[106] A difference
is that the nearest neighbor O 2p orbitals contribute more significantly to the bonding in
LiNbO2, giving a larger nearest neighbor hopping. The dz2 lobes above and below the
transition metal appear to be somewhat more pronounced in TaSe2. We also observe
changes toward such appearance in this Nb WF as the O layer height (z) is increased,
forcing the O atoms closer to the Nb layers and increasing t1. Given in Fig. 6.5 is a plot of
the difference in the magnitudes of the WFs taken for the two different z values of 0.1359
and 0.1263 (O displacements of roughly 0.1 A.) There is an increase in the density of the
dz2 lobes above and below the transition metal as well as a tilting of the O orbitals toward
the central site, as the Nb-O distance is reduced. Correlating this with the tight binding
results from the next section suggests there is an increase in the role the O atoms play in
the hopping to nearest neighbors as the Nb-O distance is decreased. The opposite effect
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 104
occurs when z = 0.1167; the O orbitals tilt to become oriented more along the z-direction,
indicating the O atoms are interacting less strongly with Nb.
6.6 Tight Binding
Γ M K Γ A L H A-5
-4
-3
-2
-1
0
1
2
3
Ene
rgy
(eV
)
Figure 6.6: (Color online) Band structure for O layer position z = 0.1167, a larger Nb-O layer separation by 0.1 A compared to the bands in Fig. 6.1. The band width is25% smaller, and band gap is significantly less, compared to the experimental structure(z = 0.1263).
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 105
Γ M K Γ A L H A-5
-4
-3
-2
-1
0
1
2
3
Ene
rgy
(eV
)
Figure 6.7: (Color online) Band structure for O layer position z = 0.1359, about0.1 A closer to the Nb layer than in experiment (Fig. 6.1). The bandwidth is 15% largerthan for the experimental structure, and the peaks at M and K in the valence band aresimilar to the band structure calculation of Li0.5NbO2 shown by Novikov et al.,[95] sug-gesting that the underlying electronic structure of LixNbO2 is determined by the positionof O layers rather than the Li concentration.
Table 6.1: Tight binding hopping parameters for three values of the O internal structural parameter z. Hopping parameters arereported in meV; bandwidths and band gaps are reported in eV. Distance between Nb and O atoms are reported in angstroms.The largest hopping coefficients are in boldface. Squeezing the O layers closer to the Nb layers (z = 0.1359) results in a dramaticincrease in t1, making it larger than t2, and a large change in t3 as well. For z = 0.1359, the band gap is indirect. The directgap at Γ is listed followed by the indirect gap in parentheses. See figure 6.7. t⊥3 is slightly refined, to make the splitting at Γagree with the DFT result better.
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 107
The valence bands seen in Fig. 6.1 are well isolated from bands above and below. These
bands contain the superconducting electrons in de-lithiated LixNbO2, so a simplified tight
binding fit using only these bands would be useful in doing studies on simple models. The
TB model to fit two bands corresponding to identical Nb dz2-O p hybridized Wannier
functions on the two Nb atoms per cell gives dispersion relations from the roots of
0 =
∣
∣
∣
∣
∣
∣
∣
ε‖k − εk ε⊥k
ε⊥k ∗ ε‖k − εk
∣
∣
∣
∣
∣
∣
∣
=⇒ εk = ε‖k± |ε⊥k | (6.2)
where ε‖k
arises from the hopping processes within a plane of atoms and ε⊥k from hopping
between planes. Within the plane of the triangular lattice, the nearest neighbor dispersion
for the six nearest neighbors is
ε‖k
= 2t
[
cos(kxa) + 2 cos
(
1
2kxa
)
cos
(√3
2kya
)]
. (6.3)
We have found that a good fit to the Nb dz2 bands requires several neighbors. Third and
fifth neighbors in the plane can be included with this form by noticing odd neighbors just
require a larger lattice constant. Second neighbors require a larger lattice constant and a
30 rotation. There are twelve fourth neighbors, which can be treated as two lattices with
separate rotations.
First and second neighbors in adjacent layers can be included with terms of the form
ε⊥k = 2coskzc
2
2t⊥1
[
cos
(
kxa
2
)
eikya
2√
3 + eikya√
3
]
+
2t⊥2
[
cos (kxa) e−
ikya√3 + e
−i2kya√3
]
. (6.4)
It is worth noting that on the zone surface (where kz = ±πc ), all terms in ε⊥k , including fur-
ther neighbors neglected from Eq. (6.4) are identically zero, resulting in exact degeneracy
of the two valence bands at these zone surfaces.
Using the Wannier basis given in the previous section, the hopping parameters can be
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 108
directly calculated (rather than fit) as matrix elements of the LDA Hamiltonian:
t|R|,m = 〈Rm|H |0m〉, m = 1, 2; (6.5a)
t⊥|R|,m,m′ = 〈Rm′|H|0m〉. (6.5b)
Using the definition of the WF along with orthogonality of the eigenfunctions, Eqs. (6.5)
reduce to
t|R|,m = N−1k
∑
kn
εkn|Ukmn|2eik·R (6.6a)
t⊥|R|,m,m′ = N−1k
∑
kn
εkn(Ukm′n)
∗Ukmne
ik·R. (6.6b)
where R is the difference in centers of the WFs. Deviation of a perfect representation
of the bands arises only from approximation of the summation in Eq. (6) or neglect of
further neighbors.
The results of the tight-binding calculation are presented in Table 6.1, as calculated
from a 12×12×6 Monkhorst-Pack grid of k-points for various values of z, along with values
for 2H-TaSe2 as given in Ref. [106]. Remarkably, the second neighbor hopping integral t2
is the dominant contribution for z ≤ 0.1263. Also notable is the strong dependence of first
and third neighbor hoppings, t1 and t3, on the Nb-O distance, which causes t1 to become
largest for z somewhat greater than 0.1263. The other parameters in the table are all
relatively independent of the Nb-O distance. The third and beyond interplane hoppings
may appear small, but they must be compared keeping in mind that there are twice as
many third and fourth interplane neighbors as first and second. Since t⊥1 ∼ 2t⊥3 ∼ 2t⊥4 ,
even these long distance hoppings are relevant. Indeed, it is not possible to get good
agreement in the band splittings at the points Γ, M, K, and the minimum between K and
Γ without five interplane hopping parameters (four for z = 0.1167).
The TB model for z = 0.1167 is strongly dominated by the second neighbor interactions
between the Nb atoms, with other hoppings roughly an order of magnitude smaller. If t2
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 109
were the only hopping parameter present, the system would consist of three disjoint sub-
lattices and thus states would be threefold degenerate (although it would not be obvious
on the standard band plot). Including t1 as a perturbation would couple them weakly and
remove degeneracies. In this case, the unit cell of the system could be taken as the second-
neighbor lattice and the Brillouin zone would be folded back so that the K(H) point in the
band plots shown is mapped onto the Γ(A) point. The result of this folding back (and the
splitting that occurs because of non-zero t1), is evident in Fig. 6.6, in that the eigenenergy
of the H point is nearly the same as for the A point. The M(L) point is mapped onto itself,
and the minimum between Γ(A) and M(L) is mapped onto the K(H) point, which results
in a band structure that looks very much like a√
3 larger triangular lattice with only
nearest neighbor hopping. Similar dominance of 2nd neighbor hopping has been observed
for 2H-TaSe2. [106] Those authors relate their small value for t1 to phase cancellation of
Wannier functions centered at first nearest neighbors, whereas in-phase contributions at
second-neighbors give a large t2 matrix element.
For z = 0.1359, we find that the nearest neighbor interaction becomes dominant. From
Table 6.1 and Fig. 6.5 we see that the increase in hybridization of the O anion p-state in
LiNbO2 as the Nb-O distance is decreased plays a significant role in first neighbor hopping,
while at the same time this hybridization is rather insignificant in second neighbor hopping.
This behavior suggests that a significant contribution to the hopping integrals comes from
within the Nb planes, perhaps through the hotdog structure noted in the previous section.
The band structure in this case looks similar to that of Li0.5NbO2, calculated by Novikov et
al.,[95] especially in the magnitude of the variations around the M and K points, relative to
the bandwidth. Their result shows some splitting of the bands occurring at the zone surface
(along the lines A-L-H-A), which may arise from the two Nb sites becoming inequivalent
due to the removal of one Li atom. If the Nb sites were kept equivalent as the Li is
removed, such as in the virtual crystal approximation, then the similarities between the
band structures suggest that the main impact on the electronic structure is captured by
changing the Fermi energy and allowing the O layers to move closer to the Nb layers,
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 110
Symmetry Phonon frequency (meV) α for TO α for LODesignation Polarization TO LO αLi αNb αO αLi αNb αO
20 B1u z 89.5 0.13 0 0.3821 B2g z 94.8 0 0.11 0.39
Table 6.2: Calculated zone-center phonon frequencies, with LO modes listed where theyoccur. See Appendix 6.7.1 for detailed descriptions of the atomic motions. All phononswith x-y polarization are doubly-degenerate, as required by the hexagonal symmetry. Thelast six columns show the calculated value of αi, defined by equation 6.7 for both TO andLO modes.
decreasing the Nb-O distance.
Another point of interest we see in Figs. 6.6 and 6.7 is a significant change in how the
valence band is situated in relation to the O 2p bands. Nb 4d states in the conduction band
around the K point are significantly lowered in energy as the Nb-O distance is decreased so
that the bandgap becomes indirect and much smaller. This change will have contributions
from both a Madelung shift and from changes arising from the altered Nb-O interaction.
6.7 Zone Center Vibrational Modes
We have calculated the 3N − 3 = 21 optical phonon eigenmodes with q = 0. To our
knowledge, experimental IR or Raman measurements have not been reported on LiNbO2.
However, we believe these results will be useful for comparison with electron-phonon cal-
culations done on the de-lithiated LixNbO2, which will be explored in future work. The
frequencies are given in Table 6.2. Mode symmetries have been obtained with the use of
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 111
SMODES, part of the Isotropy package of Stokes and Boyer.[107] Masses of the ions used
were 6.94 (Li), 92.91 (Nb), and 16.00 (O), all in amu. Also in Table 6.2 are given values
of the isotope shift α, defined by
αi = − ∂ lnω
∂ lnMi, (6.7)
where i runs over the different (types of) atoms in the crystal. The derivative was ap-
proximated with a centered difference, which, for a harmonic crystal with ω ∝ M−1/2,
introduces errors of order ( δMM )2. Using δM = 1 amu, this amounts to an error of about
2% for Li, and less for heavier elements. One would expect∑
i αi = 12 for a harmonic
mode.
The dielectric tensor for a hexagonal system such as this will be diagonal with two
distinct elements. For frequencies much higher than phonon frequencies but well below
the gap, the static dielectric constants, calculated without ionic contributions, are found
to be ǫx∞ = 10.3 and ǫz∞ = 3.78. One can then use the generalized Lyddane-Sachs-Teller
relation, ǫα0 = ǫα∞∏
m(ωαLO,m/ωαTO,m)2 (where α runs over the cartesian directions, and the
product is taken over all IR active modes with polarization in that direction) to include the
ionic contribution for the low frequency dielectric constant. Doing so, we find for LiNbO2
that ǫx0 = 15.9 and ǫz0 = 8.27. It is typical for LDA calculations to overestimate somewhat
the dielectric constants, presumably due to the band gap being underestimated.
The eigenmodes fall into groups which are easily distinguishable by frequency. In the
following discussion we include degeneracy in the numbering of modes. The first group,
modes 1-3, are low frequency modes involving displacements of massive Nb-O layers against
each other, with Li layers remaining fixed. The second group, modes 4-7, involve the lighter
Li layers, and lie around 29 meV. The third group, modes 8-17 in the 53-65 meV range,
involve motions where layers of atoms slide in the x − y plane against adjacent layers of
atoms, or groups of layers bounce against each other in the z direction. The last group
of modes at 85-95 meV involves layers nearby each other bouncing against each other in
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 112
the z direction. A more detailed discussion of the individual modes is given in Appendix
6.7.1.
6.7.1 Phonon mode descriptions
Following is a descriptive list of the atomic motions for each mode given in table 6.2,
grouped by their spectroscopic activity. Degeneracy is included in the numbering of modes.
Silent Modes
Modes 1-2: NbO2 layers slide against each other in the x-y plane.
Mode 3: NbO2 layers bounce against each other in the z direction.
Modes 4-5: LiO2 layers slide against each other in the x-y direction. Li displacements
are roughly an order of magnitude larger than O displacements, which results in the mass
of the O atoms having very little effect on the frequency of this mode.
Modes 11-12: Adjacent O layers slide against each other, with Li and Nb layers re-
maining stationary.
Mode 15: LiO2 layers bounce against each other in the z-direction. This is the one
high frequency mode whose isotope shift α gets its major contribution from the Li atoms.
Mode 20: Out-of-phase Li layers beat against adjacent O layers in the z direction.
Mode 21: Out-of-phase Nb layers beat against adjacent O layers in the z direction.
IR Active Modes
Mode 6-7: In-phase Li layers sliding against NbO2 layers. The main contribution to α
comes from the light Li atoms, with the heavy NbO2 layers making very little contribution.
This mode is also IR active, with a small LO-TO splitting ∆ω = ωLO − ωTO = 3.6 meV.
Mode 8: The TO mode consists of Li layers bouncing against NbO2 layers in the z
direction, and can be thought of as the “z-polarized version” of modes 6 and 7. When
the macroscopic electric field is included, the O displacements change sign and the mode
becomes Li-O layers bouncing against Nb layers. ∆ω = 16.7 meV.
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 113
Modes 13-14: Li and Nb are in synchronized sliding against O layers. The metal layers
are all in phase with each other, and out of phase with the O layers. This mode is IR
active, with ∆ω = 7.2 meV.
Mode 19: The transverse mode consists of LiO2 layers beating against Nb layers in
the z-direction. The Li-O layers are all in phase. This mode is IR active, and once the
macroscopic electric field is included, this mode can be characterized as having Nb layers
bounce against adjacent O layers, where the Nb layers are out of phase. ∆ω = 10.8 meV.
Raman Active Modes
Modes 9-10: Out-of-phase Li layers slide in the x-y direction against adjacent O layers.
The rest of the phonon modes from here on have the contribution to their isotope shift α
dominated by O.
Modes 16-17: Out-of-phase Nb layers slide against adjacent O layers in the x-y plane.
Mode 18: The O layers beat against each other in the z-direction. This fully sym-
metricRaman active mode is of particular interest, as it corresponds to the variation of
the internal structural parameter z. This mode has a significant impact on the electronic
structure as discussed in previous sections, so we suspect that it may have particularly
strong electron-phonon coupling in hole-doped LixNbO2.
6.7.2 Clarifying Comments
The z-polarized A2u IR active modes 8 and 19 have their eigenvectors affected by the
inclusion of the macroscopic electric field. Without the electric field, the eigenmodes can
be described as one where Nb layers bounce against Li-O units and one where Li layers
bounce against Nb-O units. The application of the electric field causes these vectors to
mix, and the descriptions change so that the latter one is replaced by a mode where Li
and Nb bounce against O layers. The distinction between these two modes is the relative
sign for Li and Nb displacements (negative for the former, positive for the latter.) It is
not unusual to have modes mix to give different LO modes, as the non-analytic part of
Table 6.3: Calculated Born effective charges for LiNbO2, compared with those calculatedfor NaCoO2 by Li, Yang, Hou and Zhu. [108] Note the disparity in the effective chargesfor the alkali metals and the transition metal in the z-direction. For O however, Z∗ is notfar from isotropic.
the dynamical matrix which incorporates the electric field can have eigenvectors differing
from the analytic part of the dynamical matrix, resulting in a mixing of modes of the same
symmetry. This is discussed in some detail by Gonze and Lee.[17] This complicates the
association of LO modes with their corresponding TO mode, so we have chosen to pair
the LO and TO modes such that ωLO > ωTO.
6.8 Effective charges
Born effective charge calculations have been carried out as described by Gonze and Lee.[17]
The calculation was done using a planewave cutoff of 60 Ha and 200 k-points in the
Brillouin zone to give effective charge tensors Z∗ for each atom which are diagonal, and
obey Z∗xx = Z∗
yy, as required by symmetry for a hexagonal crystal. Results are given in
Table 6.3 for the relaxed atomic structure. The effective charges of NaCoO2 have been
included in this table to illustrate the strong similarity between these two systems, which
are similarly layered transition metal oxides but otherwise their electronic structures are
quite different.
The effective charges of LiNbO2 are rather different than those reported by Veiten and
Ghosez[84] for LiNbO3. There are structural and chemical differences responsible for this;
the structure of LiNbO3 has Nb coordinated in a distorted octahedral environment and thus
has low symmetry. Perhaps more importantly, the formal charges of Nb are very different:
Nb3+ in LiNbO2, but Nb5+ in LiNbO3, leaving the formal electronic configuration of Nb
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 115
in LiNbO3 as d0. The effective charges of such ions are known to sometimes be rather
large, as is the case in LiNbO3.
Z∗xx(Li) is close to the formal charge of Li indicating ionic type response for in-plane
motion. In fact, all ions have effective charge closer to the formal value for in-plane
displacement than for z displacement. The charge tensor for Li shows similar anisotropy
to that of Li in LiBC,[109] where it was inferred that Li is involved chemically in coupling
between B-C layers. Similarly here, it can be expected that Li has chemical involvement
in interlayer coupling between the electron rich Nb-O layers more so than Na does in
NaxCoO2. This involvement is not strong enough to inhibit de-intercalation, however.
The smaller than formal charges for Nb indicates substantial covalent character to the
bonding, which is especially strong in the z direction. This charge ‘renormalization’ may
be related to the strong change in band structure due to change of the distance between
the Nb and O layers, discussed in earlier sections.
6.9 Summary
In this work we have examined the electronic structure, the Wannier functions of the
valence bands and their tight-binding parametrization, and several vibrational properties
of the quasi-two-dimensional material LiNbO2. These basic properties are important in
providing the basis for understanding the behavior of the system when it is hole-doped by
de-intercalation of Li.
LixNbO2 seems to be a candidate for an insulator-to-superconductor transition, as
it undergoes an insulator-to-metal transition and becomes superconducting, but doping
studies are not systematic enough yet to shed light on this possibility. Phase space con-
siderations in two-dimensional materials lead to superconductivity that is independent of
doping level, at least when the Fermi level is not far from a band edge.[100] This property
supports the possibility of an insulator-superconductor transition with doping, i.e. as soon
as it becomes metallic it is superconducting.
CHAPTER 6. WANNIER FUNCTIONS IN LINBO2 116
The Wannier function, centered on and based on the occupied Nb dz2 orbital, illustrates
graphically not only how Nb bonds with neighboring O ions, but also that it couples with
oxygens in adjoining layers. Inspection of the k-space wavefunction suggests there are
direct Nb-Nb interactions within the dz2 band as well as indirect coupling through O ions.
A tight-binding representation of the band, obtained by direct calculation (rather than
fitting) using the Wannier function, reveals that several neighbors both in-plane and inter-
plane are necessary to reproduce the dispersion. To model the interlayer interaction pre-
cisely requires four to five interlayer hopping parameters. For the experimental structure
intralayer hopping is dominated by 2nd neighbor hopping, but the strength of both 1st and
3rd neighbor hopping is strongly modulated by varying the Nb-O distance. This electron-
lattice coupling provides one possibility for the superconducting pairing mechanism in
LixNbO2.
Comparison with previous calculations of the band structure of Li0.5NbO2 suggests
that the electronic structure of LixNbO2 can be modeled in virtual crystal fashion, or
perhaps even in rigid band fashion if that is convenient. The Nb dz2 band is remarkably
well isolated from neighboring bands by Madelung shifts and crystal field splitting. We
suspect that the overriding importance of this system will result from its unique standing
as a single band, triangular lattice system for which sophisticated studies on simple models
may be directly comparable to experimental data.
6.10 Acknowledgments
The work described in this chapter was done in collaboration with W. E. Pickett. We
have benefited from communication with K. E. Andersen and A. Simon. This work was
supported by National Science Foundation Grant DMR-0421810.
117
Chapter 7
Gruneisen Parameters and
Equation of State
The work in this chapter was done in collaboration with Phil Sterne from LLNL.
7.1 Introduction
Detailed information about the equation of state of materials is of considerable scientific
and technological importance. New research facilities, such as the National Ignition Facility
(NIF) at Lawrence Livermore National Labs (LLNL) have been recently developed which
are capable of testing materials under extreme pressure and temperature conditions.
Many elements which are thought to be simple nearly free-electron-gas metals in am-
bient conditions have recently attracted a great deal of scientific interest due to reports
unusual properties appearing under pressure. It was theoretically predicted that Li would
be superconducting under pressure,[110] and later it was experimentally discovered to
be superconducting up to 20 K around 48 GPa,[111] which at the time was the highest
known superconducting transition for an element. This has recently been explained as
BCS-type phonon mediated superconductivity with strong electron-phonon coupling.[112]
Since then, Ca has been reported to be superconducting with TC = 25 K[113, 114] under
CHAPTER 7. GRUNEISEN PARAMETERS AND EQUATION OF STATE 118
pressure, which is currently the highest known TC for an element. Metallic Y also shows
a fairly high superconducting transition at 17K under pressure,[115] which has also been
explained as strong electron-phonon coupling in a recent theoretical study.[116]
Aside from superconductivity, several other surprises have been discovered in the phase
diagrams of elements under pressure. Above 32 GPa, Ca has been reported to exist in
a simple cubic phase[117] which has been theoretically verified.[118] The existence of the
simple cubic phase, especially under pressure, is quite surprising due to the low coordina-
tion and the tendency for group I and II metals to favor close-packed structures already
at low pressures. The melting transition in the Na phase diagram has been seen to rise
from 370K up to 1000K at ∼35 GPa and back down to 300K at 120 GPa.[119] Although
it is not unusual to have a melt curve with a negative slope, the unusual characteristic
seen in sodium is that the melting temperature drops 700K over three solid phases (bcc,
fcc and cI16) of Na. The discovery of this transition has lead to considerable theoretical
interest,[120, 121, 122, 123] particularly in the use of first principles molecular dynamics
(FPMD) to model the melting.[120, 121, 122] In other group I metals (K and Cs), a bcc
to fcc transition is observed, as well as decreases in the melting temperatures as solid-solid
phase boundaries are approached, but these drops are relatively small, the largest being
100 K in fcc Cs.[124]
There are several elements which exhibit so-called correlated behavior. Lanthanides
such as Ce, Pr and Gd have drawn considerable attention due to the presence of anoma-
lous volume collapses when placed under pressure.[125] These materials require methods
which can treat the f electrons on a more sophisticated footing than the local density
approximation (LDA), such as Dynamical Mean Field Theory.[126, 127]
Considerable effort has been expended in attempting to map the phase diagram of
carbon.[128, 129] Experimental measurements of the diamond melt curve are quite dif-
ficult, considering the very high temperatures required to melt diamond and that these
types of experiments are usually done in a diamond anvil cell. The behavior of carbon un-
der extreme pressures (which cannot be obtained via traditional laboratory methods) is of
CHAPTER 7. GRUNEISEN PARAMETERS AND EQUATION OF STATE 119
considerable importance in geology and astrophysics. There are however significant imped-
iments to this understanding; different theoretical models have difficulties agreeing with
experiment or with other theoretical calculations on the diamond melt curve.[128, 129, 130]
Here we present a Density Functional Theory (DFT) study on three elemental systems,
Al, Na and diamond-phase C. Our interest here is to demonstrate the suitability of this
method for equation of state research, and to examine some simple rules for calculating
melt curves from first-principles.
7.2 Theoretical Background
7.2.1 Gruneisen Parameter
The Helmholtz free energy F for a non-interacting phonon gas is given by
F (V, T ) =1
VBZ
∑
s
∫
BZ
[
~ωq,s
2− kT ln
(
1 − e−β~ωq,s
)
]
d3q, (7.1)
where q is a vector within the first Brillouin zone, and s is a band index. We also define
the dimensionless thermal Gruneisen parameter γ as
γ =V
CV
(
∂2F
∂T∂V
)
=αV B
CV, (7.2)
where α is the coefficient of thermal expansion, B is the isothermal bulk modulus, V is
the volume, and CV is the heat capacity. It is straightforward that in the quasi-harmonic
approximation, in which the phonon frequencies ωq,s depend on volume and not temper-
ature,
γ = 〈γq,s〉CV(7.3)
where γq,s is the mode Gruneisen parameter
γq,s = −∂ lnωq,s
∂ lnV, (7.4)
CHAPTER 7. GRUNEISEN PARAMETERS AND EQUATION OF STATE 120
and the subscripted CV indicates that the average over each mode is weighted by its
contribution to the specific heat. In the quasi-harmonic approximation we have γq,s =
γq,s(V ), while the averaging against specific heat gives γ its temperature dependence.
At temperatures T ≫ ΘD, all modes contribute equally to specific heat, so γ will be
independent of temperature. For the systems studied here, this holds true even for T < ΘD,
down to about T . 100K.
7.2.2 Melting Rules
The original theory of melting was proposed by Lindemann[131] where it was proposed
that melting happens once the mean square deviation of atoms from their lattice positions
reached some threshold value. Once the phonon frequencies ωq,s and γ(V ) are known, one
can calculate the melting temperature of the material using the Lindemann criteria [132],
d lnTmd lnV
= −2
[
γ(V ) − 1
3
]
. (7.5)
Integrating (7.5) and defining γ = −d ln ΘD
d lnV (as is commonly done) gives
Tm = AΘ2DV
2/3 (7.6)
where A is an integration constant, which can be calculated in terms of physical constants
and characteristics of the material under study,[132] however for this work we consider A
to be a fitting parameter.
The author of Ref. [132] makes arguments based on the change in entropy that occurs
during melting in order to effectively replace the characteristic temperature and volume
in (7.6) to get
Tm = Aω20V
2/3lm (7.7)
where the Debye temperature has been replaced by ω0, the zero-th phonon moment, and
the volume used is the volume of the liquid phase at melting. The phonon moments for
CHAPTER 7. GRUNEISEN PARAMETERS AND EQUATION OF STATE 121
n > −3 are defined as follows,
ωn =
[
n+ 3
3
∫ ∞
0ωng(ω)dω
]1/n
, n 6= 0 (7.8)
ω0 = exp
[
1
3+
∫ ∞
0lnω · g(ω)dω
]
. (7.9)
where g(ω) is the normalized phonon density of states. (A moment for n = −3 may also be
defined, but that is not relevant for our purposes here.) The numerical factors are chosen
for ease of comparison with a Debye model. For a normalized Debye density of states,
(g(ω) = 3ω2/ω3D for ω ≤ ωD, 0 otherwise) it is straightforward to show that all ωn are
equal to the Debye frequency. [132]
Certain phonon moments have specific physical interpretations, in particular the mo-
ment ω−2 scales the classical nuclear mean square displacement. [132] Considering the
original spirit of the Lindemann criteria, that melting occurs when the mean square dis-
placement of atoms reaches some threshold value, we suggest an alternate melting rule
Tm = Aω2−2V
2/3. (7.10)
We investigate the quality of these melting rules in the next section. Since we do not have
a good way to estimate Vlm from V , we simply use the volume of the solid at the melting
temperature in (7.7) in the next section.
7.3 Method
All calculations were performed within Density Functional Theory (DFT) with a planewave
basis, using the Abinit code versions 4.4.4 and 4.6.5.[103] Pseudopotentials used were
Troullier-Martins type [18], with three electrons for Al, one electron for Na and four
electrons for C.
Phonon frequencies were calculated on a Monkhorst-Pack grid of q-points in the irre-
CHAPTER 7. GRUNEISEN PARAMETERS AND EQUATION OF STATE 122
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Volume (%)
0
0.5
1
1.5
2
Gru
neis
en P
aram
eter
(γ)
Al
bcc Na
Diamond
Figure 7.1: (Color online) Calculated Gruneisen parameter of the materials selected forthis study.
ducible Brillouin zone. The q grid was unshifted so that Γ and zone boundary points were
included. This is the most convenient approach to calculate dynamical matrices and to
insure the acoustic sum rule is obeyed. Inclusion of Γ point phonons presents problems in
calculating the numerical derivatives in (7.4), so the phonon frequencies were interpolated
onto a shifted grid which does not include Γ for calculations of γ(V ).
For Al, the number of k-points used was increased as the volume was decreased, in order
to maintain a high resolution of the Fermi surface. This same approach was attempted
with Na, but we found it produced noticeable ‘kinks’ in the results around the volume
where the number of k-points was changed.
7.4 Materials
We chose three elemental materials to test the theoretical methods presented in the previ-
ous sections. The first of these systems was Al, which was chosen because it is well known
that Al is a rather free-electron like metal and treated well within present formulations of
DFT. The Al pseudopotential used is a norm-conserving Troullier-Martins type [18], with
three (3) valence electrons.
Sodium was chosen as the second system because the recent discovery of its unusual
CHAPTER 7. GRUNEISEN PARAMETERS AND EQUATION OF STATE 123
melt curve [119] provides an opportunity to test the various melting rules from section
7.2.2 against each other on a system which is known to have a negative melt slope for an
appreciable pressure range.
The diamond phase of carbon was chosen as the third system to study because it is an
elemental system for which there are relatively few phase boundaries, as well as it is distinct
from Na and Al in that it is an insulator, and there are two atoms per unit cell so optical
modes must be considered. The melting transition of diamond is also of considerable
interest in astrophysical research in understanding diamond-methane transitions which
may occur in the core of gas giants. Calculations for diamond were not carried out below
60% of the equilibrium volume, as this has already reached a pressure of 1000 GPa.
We have calculated the volume dependent Gruneisen parameter for these three elemen-
tal systems over a large range of pressures from first principles. Shown in Fig. 7.1 is a plot
of the Gruneisen parameter for Na, Al, and the diamond phase of C. The tendency for γ
to go negative at large compressions (V . 35%) is due to lattice instabilities from phonon
modes going to zero frequency. In the DFT calculations, these instabilities eventually
cause the phonon frequencies to become imaginary, so near these instabilities it becomes
difficult to use finite differences to calculate the derivatives of ωq to evaluate (7.4).
7.4.1 Aluminum
Shown in Fig. 7.3a are plots of the phonon moments ωn as defined by (7.8) for Al. For a
given volume, ωn for n > −2 is rather independent of n, indicating Al is very Debye-like
even up to high pressure. The ‘flatness’ of these curves indicate that melt rules based on
ω−2 and ω0 will not give significantly differing results. Thus, in Fig. 7.2 we only show
the melt curve generated from the Debye temperature. The agreement with experimental
data points[133] as well with other theoretical calculations based on a more sophisticated
approach involving calculation of anharmonic contributions and the free energy of the
liquid phase[134] is quite good.
CHAPTER 7. GRUNEISEN PARAMETERS AND EQUATION OF STATE 124
0 10 20 30 40 50 60 70 80 90 100
Pressure (GPa)
1000
2000
3000
4000
Mel
ting
Tem
pera
ture
(K
)
Figure 7.2: (Color online) Plot of melt curve of aluminum as calculated from (7.6), com-pared with experimental points, from ref. [133]. The agreement is very good.
7.4.2 Sodium
Calculations for Na were done both in the bcc and fcc phases. To test whether the
pressures calculated using the pseudopotential method would be accurate due to the loss
of interaction between the 2p semi-core states, total energies were calculated using the
full-potential local orbital (FPLO) code. The total energy curves generated by Abinit
and FPLO were very similar, even at high pressures.
Sodium-BCC phase
Shown in Fig. 7.4 are the melt curves as calculated by (7.6), (7.7) and (7.10), compared
to experiment[119] and first principles molecular dynamics (FPMD).[121] The Lindemann
rule based on ω−2 gives significantly better agreement with experiment than the rules
based off ω0 or θD in the region where the melt curve has a negative slope (P & 20 GPa),
as well as the pressure at which the maximum occurs. The agreement with experiment
in this phase is at least as good as the agreement between the FPMD calculation and
experiment.
The source of the improvement upon using ω−2 is clear. Under high compression, the
bcc to fcc lattice instability plays a role in softening phonons, to which ω−2 is much more
CHAPTER 7. GRUNEISEN PARAMETERS AND EQUATION OF STATE 125
0
500
1000
1500
2000
2500
-3 -2 -1 0 1 2 3 4 5 6 7 8 9 100
100
200
300
400
500
600
700
(a) Al
(b) Na
1.00
0.75
0.50
0.35
0.28
1.00
0.75
0.50
0.25
Figure 7.3: (Color online) Shown here are the phonon moments ωn plotted vs. n (seeeqn. 7.8) for (a) Al and (b) Na at selected volume percentages. A flat line indicates thefrequency distribution is very Debye-like. Al phonons are very Debye-like, even at highcompressions, but Na phonons become less Debye-like under pressure.
sensitive to than ω0. In Fig. 7.3 we see this causes ω−2 to drop much more than ω0 at high
compressions. At lower pressures, the ωn vs. n curve is rather flat, so the two moments of
interest here have very similar values. This gives us a very simple physical picture of why
there is a maximum in the melt curve and a drop in the melting temperature as pressure
is increased – the bcc lattice is being driven to instability where the fcc lattice will become
the stable phase. Intuitively speaking, as the bcc lattice becomes more unstable, atoms
are ‘less bound’ to their crystallographic positions so less thermal energy is required for
melting to occur. The lowest frequency modes dominate this displacement, because they
are easiest to occupy. To first order in ω/T , the occupation of a mode with frequency ω is
T/ω, so the root mean square displacement is roughly proportional to√T/ω. As a mode
softens under increasing pressure, its displacements get larger at constant temperature
allowing melting to occur at a lower temperature. This effect is captured rather well here
by the melting rule based on ω−2.
Sodium-FCC phase
For the fcc phase of Na, there are only three experimental points from ref. [119], and
the DFT-GGA calculations are not harmonically stable near 100 GPa, so we are unable
CHAPTER 7. GRUNEISEN PARAMETERS AND EQUATION OF STATE 126
20 40 60 80 100Pressure (GPa)
300
400
500
600
700
800
900
1000
Mel
ting
Tem
pera
ture
(K
)
Melt rule from θD
Melt rule from ω0
Melt rule from ω-2
bcc fcc
Figure 7.4: (Color online) Plot of the phase diagram sodium, as calculated from (7.6),(7.7) and (7.10), compared with experiment and ab initio molecular dynamics. Solid linesindicate melt curves calculated for bcc Na, and dashed lines indicate calculations performedfor fcc Na. Solid symbols indicate experimental data, open symbols indicate moleculardynamics data. Solid squares are bcc phase melting temperatures, solid diamonds are fccphase melting temperatures.[119] Solid circles are bcc melting temperatures from ref. [124].Up triangles indicate simultations of solid phase, and down triangles indicate simulationsof liquid phase.[121] The vertical dashed line indicates the bcc-fcc transition given by ourcalculation.
to draw very useful conclusions about the agreement of theory with experiment in this
region. The fcc melting curves in Fig. 7.4 are stopped below 100 GPa due to these
instabilities. There is a systemic underestimation of the stability of a phase with respect
to pressure. Experimentally the bcc → fcc transition pressure should be 65-70 GPa, but
our calculations predict this transition to occur at 52 GPa. We did not investigate the
fcc-cI16 transition.
In the FPMD calculations it was found that there is a transition in the liquid above
about 65 GPa where a pseudogap forms in the density of states for the liquid phase that
does not appear in the fcc phase, and the coordination of atoms in the liquid becomes
much more like that of the high pressure cI16 phase. It is most likely that the failure of
the Lindemann melt rules in the fcc phase is related to the fact that this sort of structure
in the liquid phase is ignored. Interestingly, the melt curve from (7.10) tracks the melting
temperature of the FPMD calculations well into the fcc region.
If we compare the three curves from the melt rules, we see that the curves from ω0
CHAPTER 7. GRUNEISEN PARAMETERS AND EQUATION OF STATE 127
and θD look very similar; there is little to be concluded from comparing these two curves.
The curve from ω−2 is quite unrealistic. It is somewhat surprising the the ω−2 curve
performs unrealistically when it gives very good agreement with the bcc phase. It is a
clear consequence of the results presented in Fig. 7.3 where we see that ωn is rather
insensitive to the phase transition from bcc to fcc, but the experimental melt curve shows
a change in slope at the phase transition. In fact, the integration constant in any of the
calculated fcc melt curves can be adjusted so that they lie on top of the calculated bcc
melt curves. This actually would agree quite well with the FPMD calculations up to
about 90 GPa. Above this, anharmonic effects may play a significant role in stabilizing
the system, and if so, then this would have an impact on the melt curve and certainly
cause the Lindemann criteria to be inadequate in this region.
7.4.3 Diamond
Shown in Fig. 7.5 is the melting curve for diamond calculated from (7.7) and (7.10) com-
pared to results from a variety of other theoretical methods.[129, 130, 128] For each (7.7)
and (7.10), two different values of the arbitrary constant are chosen to show agreement
with calculations from ref. [130] (lower curves) and ref. [129] (upper curves). The agree-
ment of the upper ω−2 melt curve with results from the ab initio molecular dynamics
calculations[129] is rather good, although our melt curve does not show a maximum in
this pressure range. The melt curve from ω0 does not agree quite so well. Both the lower
curves’ agreement with calculations based on Brenner’s bond order potential[130], is quite
good up to 100 GPa, and the ω−2 melt curve performs slightly better in the 100-200 GPa
range.
The agreement of our calculations of the Lindemann model with the other theoreti-
cal methods presented here is not significantly worse than the agreement between those
other theoretical models. Discrepancies in the diamond melting curve are likely caused
by several factors. A significant source of error is likely the anharmonicity from phonon-
phonon interactions which becomes stronger at higher temperatures, which is the break
CHAPTER 7. GRUNEISEN PARAMETERS AND EQUATION OF STATE 128
0 200 400 600 800 1000Pressure (GPa)
4000
5000
6000
7000
8000
9000
10000
Mel
t Tem
pera
ture
(K
)
Figure 7.5: (Color online) Shown here is the melt curve calculated from ω0 (black, solidsquares) and ω−2 (blue, solid circles) compared with results from theoretical calculationsin refs. [129] (green, open diamonds), [128] (orange, open circles) and [130] (red, opensquares). Melt curves calculated in this study are shown for two different values of theintegration constant for comparison with values from refs. [129] and [130].
down of the quasi-harmonic assumption. Also diamond likely undergoes an insulator to
metal transition as it melts, further complicating things. Additionally, the liquid phase of
diamond is unusually complex, with some groups speculating about the possible existance
of a liquid-liquid phase change.[130]
7.5 Conclusion
We have seen here that DFT can provide an accurate equation of state for materials un-
der a large range of pressures and temperatures. We have presented a method based on
the Lindemann criteria for calculating the melting curve for simple elemental materials
with only a single adjustable parameter, and we have seen here that this method can
provide a reasonably accurate model for melting over a range of pressures. Unlike other
approaches for calculating the melt curve, all calculations are based on T = 0 properties of
the crystalline solids; no calculation of the thermodynamic properties of the liquid phase
is necessary. For fcc-Al and bcc-Na, the Lindemann model provides excellent agreement
with experimental melt curves, so we conclude that for these materials, the melting tem-
perature’s pressure dependence is mostly based on characteristics of the solid phase of the
material.
CHAPTER 7. GRUNEISEN PARAMETERS AND EQUATION OF STATE 129
In bcc-Na, we find that the accuracy of the Lindemann model can be improved by
choosing a characteristic temperature which is directly related to the mean square atomic
displacements. It is seen that the overturn in the melt curve is driven by phonons which
are soft at high pressures, so they have high populations at melting temperatures and thus
large atomic displacements, causing the melt curve to drop as these phonon frequencies
soften. For fcc-Na the melt curve cannot be adequately modeled with the Lindemann
model using the same characteristic temperature as is used for bcc-Na. It is likely that
anharmonic effects play a significant role in this region which are not considered in this
method.
For diamond, the melt curve gives some results which compare reasonably well with
other theoretical calculations. Molecular dynamics calculations on carbon indicate that
the liquid has unusual properties; atomic coordinations ranging from 2-fold to 5-fold are
reported[130, 135] and a variety of bonds (sp and sp3) are present.[135] These unusual
characteristics of the liquid phase probably have an impact on the melting behavior which is
not captured in this model. It is worth noting that there is significant disagreement between
various theoretical methods on the melting curve of diamond; the approach presented here
does little to help distinguish the previous calculations.
130
Chapter 8
DMFT applied to Yb Valence
Transition
The work described in this chapter was done in collaboration with Andy K. McMahan
from LLNL and Warren E. Pickett.
8.1 Introduction
It is well known that the lanthanides Yb and Eu behave rather differently from the other
lanthanides in their elemental states, due to their unusual valence states. The majority of
the lanthanide series is trivalent, however for Yb and Eu the 3+ valence state has a single
hole in the f -subshell, and it can be favorable to fill this hole with a valence electron. This
results in a number of anomalous properties, such as a larger molar volume (as compared
to the trend the rest of the lanthanide series follows), and a lower bulk modulous.[136]
They also exhibit anomalous behaviors in pressure or temperature dependent properties;
for example Yb has a thermal expansion coefficient that is three times larger than most
other lanthanides. [137] Yb exhibits a well known valence transition under pressure, where
an f electron is gradually promoted to the valence band from 0-34 GPa. Equation of state
fits based on a divalent and trivalent Yb ion do well at low and high pressures, respectively,
CHAPTER 8. DMFT APPLIED TO YB VALENCE TRANSITION 131
however the mixed-valent region between these pressures is difficult to fit well.[136]
There are a variety of ab initio methods developed to model the correlated behavior of
electrons in systems involving ions such as Yb. The most commonly used method, LDA+U
has had a great deal of success dealing with transition metals systems. It is somewhat
less well suited to to the lanthanides, particularly Yb, as Yb’s f -shell is nearly fully filled
in LDA. In this paper, we examine the pressure and temperature dependence of the Yb
valence transition using the LDA+DMFT method, which introduces an orbital dependent
potential (similarly to LDA+U) but allows for that potential to be frequency dependent,
in contrast to LDA+U. DMFT requires the choice of a impurity solver; here we compare
and contrast the results from two impurity solvers, the computationally efficient Hubbard-I
(HI) solver, and the highly accurate Quantum Monte Carlo (QMC) solver.
8.2 Calculations
Yb goes through several phase transitions under pressure, including a re-entrant fcc phase.
Experimental evidence suggests that the valence transition in Yb is relatively insensitive to
the phase, so all calculations performed in this work are done in the fcc phase. This is not
unreasonable, since we believe the important physics is occuring in the highly localized
f -orbitals, which play very little role in bonding, and the phases at both low and high
pressure have high coordination numbers.
8.2.1 DFT-LDA calculations
The LDA/GGA equation of state calculations were done in the full potential LAPW scheme
using the publicly available Exciting code. [138] LDA calculations used as the starting
point of LDA+DMFT calculations were done using the LMTO-ASA approximation and
FPLO. [22]
CHAPTER 8. DMFT APPLIED TO YB VALENCE TRANSITION 132
8.2.2 All electron LDA+DMFT
In many DMFT calculations, the Hilbert space is downfolded to only the correlated bands
in order to reduce the computation time required for expensive impurity solvers. We do
not take this approach here; indeed since the valence transition of Yb is of principal interest
here, it would be inappropriate to downfold the valence orbitals out of the basis. Thus,
the LDA+DMFT calculations presented here retain the 6s, 6p, and 5d valence orbitals
in the crystal, bath, and impurity Green’s functions. We term this method “all-Electron
LDA+DMFT” since no downfolding has been done. The self-energy is taken only to be
non-zero on the f -orbitals, and it is composed of only two distinct (frequency-dependent)
values, corresponding to the two relativistic j = 5/2 and j = 7/2 states.
The LDA calculation is done with an LMTO basis in an atomic sphere approximation.
Because of the close-packed nature in the fcc phase this should be an adequate calculation.
Two different impurity solvers are used in DMFT, the Hubbard I (HI) atomic solver and
the Hirsch-Fye Quantum Monte Carlo (QMC) solver [74]. The QMC solver is highly
accurate, but very computationally expensive. CPU time for QMC goes as O(β3) where
β = T−1 is the inverse temperature, giving a practical lower limit on temperature of about
630K in our calculations. The only systematic source of error in the QMC algorithm is
the result of the discretization of imaginary time into L ‘time slices.’ We will examine the
results for various values of L to get an idea of the magnitude of this error. Hubbard I
treats the impurity as an isolated atom, neglecting the DMFT bath function, so all physics
related to the Kondo effect are lost in this approximation.
The LDA+DMFT calculations presented here are not done with charge self-consistency.
Performing a charge self-consistent calculation is a significant challenge for DMFT, par-
ticularly if the Hilbert space is downfolded, or computationally expensive impurity solvers
are used. For Yb, charge self-consistency is likely to be important, because DMFT can
significantly modify the orbital occupations. Charge self-consistent results are currently
being explored and will be presented in a future publication.
CHAPTER 8. DMFT APPLIED TO YB VALENCE TRANSITION 133
2.5 3 3.5 4 4.5
rws
5.5
6
6.5
7
7.5
8
Ene
rgy
(eV
)
Uε
f / -13
Figure 8.1: Plot of LDA+DMFT parameters U and εf used. −εf/n0 (see text) is plottedfor comparison with U .
8.2.3 Hubbard Parameters
In doing an all-electron LDA+DMFT calculation, since the interaction is applied only
to the interacting states a double-counting energy must be subtracted because the LDA
Hamiltonians already contain a mean-field description of the interaction. The approach
used here is to simply modify the Hamiltonian used in the DMFT: Hk → Hk + Σ(ω)− εf ,
where the self-energy Σ(ω) and the double counting εf are applied along the diagonal of
the f -orbital block. The double counting parameter for each volume is calculated from
εf = −W14
(7 − n0) + U(
n0 − 12
)
+ εLDAf (8.1)
where W is the f -state bandwidth and n0 is the nominal f -electron count, 13 in the case of
Yb and εLDAf is the center of the f bands in LDA. This approach has been used previously
by McMahan et al. [139]. The value of U is computed via constrained LDA calculations
[140]. The values of U and εf are shown in Fig. 8.1.
CHAPTER 8. DMFT APPLIED TO YB VALENCE TRANSITION 134
3 3.5 4
rws
(Bohr)
0
50
100
150
200
Pres
sure
(G
Pa)
LDAGGA (PBE)PBEsolLDA+DMFT(HI)Experiment
3 3.2 3.4 3.6 3.8 40
10
20
30
40
Figure 8.2: Plot of equation of state for Yb from experiment [141], various DFT function-als, and DMFT(HI). The inset shows a blow-up of the pressure region where the valencetransition occurs. Even though LDA+DMFT(HI) shows a spurious volume collapse tran-sition at 5 GPa, its predicted equation of state is much improved over LDA or GGA.
8.3 Equation of State
Results presented from here will generally be in terms of rws, related to the volume and
lattice constant by V = 43πr
3ws = 1
4a3. Shown in Fig. 8.2 is the calculated pressure-volume
curve LDA, GGA and LDA+DMFT(HI) compared to experimental values [141]. LDA+U
results are not shown as they do not have a significant impact. This is largely due to
the fact that the f orbitals are nearly fully filled, so LDA+U does not have a significant
impact. LDA produces the worst equation of state (EOS), with VLDA(P = 0) = 0.75V0,
where V0 is the experimental equilibrium volume. The two GGA functionals used provide
some improvement of the V (P = 0) prediction, but the overall shape of the P -V curve
is unchanged and still unsatisfactory. The LDA equilibrium volume might be improved
if the f electrons are frozen into the core and a divalent ion is forced in the calculation.
Even if this approach provided reasonable accuracy for the LDA V0 predcition, it would
be unsatisfactory for calculating the equation of state over even a small pressure range
CHAPTER 8. DMFT APPLIED TO YB VALENCE TRANSITION 135
and would likely require adjustments based on empirical information about the valence
transition.
LDA+DMFT(HI) shows a much improved equation of state. Equilibrium volume is
roughly at rws = 3.97, which is a slight over-binding of the unit cell from experiment,
where the equilibrium volume is at rws = 4.05. This is about a 6% smaller volume than
experiment. Seen from the inset of Fig. 8.2, the error in the region near V = V0 is
relatively small, but since divalent Yb is fairly soft, P -V curve in the divalent region is
not terribly steep, so this results in a seemingly large error in equilibrium volume. The
major discrepancy in the LDA+DMFT(HI) equation of state is that it predicts a spurious
volume collapse at around 5 GPa. This is most likely caused by DMFT(HI) producing
the valence transition much faster than seen in experiment, discussed further in the next
section. This spurious volume collapse could also be related to structural phase transitions
to bcc and hcp that occur above 4 GPa, along with the fact that the fcc phase is re-entrant,
but we believe this is unlikely since the intermediate phases still have a high coordination
number. At rws < 3.75, the DMFT(HI) EOS predicts lower pressures than experiment,
down to approximately rws = 3.2, where the valence transition is nearly complete. Below
that volume the theoretical and experimental P -V curves agree quite well.
8.4 Orbital occupation
Shown in Fig. 8.3 is the f -orbital occupation nf vs. rws at T = 630 K for both impurity
solvers. DMFT(QMC) at L = 80 provides excellent agreement with experiment. There
may be significant error from the imaginary time discretization, as seen by the differences
in L = 80 and L = 160. In fact, the presence of this error may make the agreement with
experiment better at L = 80 than at higher L. This is most likely due to the lack of charge
self-consistency in the LDA+DMFT method. The removal of electrons by DMFT should
lower the 4f potential in LDA, which will attact electrons back into the 4f orbitals in a
charge self-consistent cycle, thus we should expect that an accurate LDA+DMFT(QMC)
CHAPTER 8. DMFT APPLIED TO YB VALENCE TRANSITION 136
3 3.2 3.4 3.6 3.8 4rws (Bohr)
13
13.2
13.4
13.6
13.8
14
n f
DMFT(HI)DMFT(QMC) L = 80DMFT(QMC) L = 112DMFT(QMC) L = 160XAESRIXS
Figure 8.3: Plot of Yb f occupation vs. rws as calculated by LDA+DMFT with HIand QMC impurity solvers, compared to experimental data. X-ray absorption edge spec-troscopy (XAES) data was taken from Ref. [142], and resonant inelastic x-ray scattering(RIXS) data taken from Ref. [143]. The DMFT calculations are performed at T = 630K, and QMC calculations were done with three different imaginary time discretizations.The L = 112 and L = 160 QMC calculations are much more difficult to converge than atL = 80.
CHAPTER 8. DMFT APPLIED TO YB VALENCE TRANSITION 137
3.2 3.4 3.6 3.8 4rws
(Bohr)
13
13.2
13.4
13.6
13.8
14
n f
HI T = 16 KHI T = 630 KHI T = 1580 KQMC T = 630 KQMC T = 1580 KXAES
Figure 8.4: Plot of Yb f occupation vs. rws at temperatures 16 K, 630 K, and 1580 K.Room temperature experimental data is shown for comparison.
procedure without charge self-consistency would underestimate nf . The calculations at
larger L values have proved difficult to converge, for reasons discussed in Sec. 8.4.1.
DMFT(HI) provides the valence transition as expected, however there is a slight overes-
timation of the tendency for the hole to appear in the f states, as compared to experiment
and DMFT(QMC). This is not too surprising, as under pressure one would expect a hole in
the f states to be only partially localized, but DMFT(HI) does not allow any delocalization
of the hole to occur. The DMFT(QMC) results closely match experiment over the entire
pressure range of interest, except for the single point at rws = 3.13. There is considerable
noise in the experimental data taken at this pressure (34 GPa) in Ref. [142], and it doesn’t
agree well with the trend in the RIXS data, so we conclude that this disagreement is not
significant.
Orbital occupation temperature dependence
The calculated nf values show a significant temperature dependence in both DMFT(HI)
and DMFT(QMC). The temperature dependent nf values are shown in Fig. 8.4. The
CHAPTER 8. DMFT APPLIED TO YB VALENCE TRANSITION 138
lowest temperature accessible in DMFT(QMC) is 630K, and its agreement with room
temperature data very satisfactory, although it is showing that the transition is not fully
completed at 630 K. At 1580K the calculated nf is lower at all volumes in than at 630 K,
suggesting that the f13 state is favored at high temperature. Both QMC and HI agree on
this trend. Also consistent with this trend, and shown in Fig. 8.4 is the nf from DMFT(HI)
at T = 16 K. This temperature is inaccessible in the QMC Hirsch-Fye algorithm, but we
expect at least the qualitative trend to be the same in QMC and HI. At T = 16 K, the
f13 state is less favored relative to the f14 state.
In DMFT(HI) the transition to the f13 state is never fully realized at very low tem-
perature, although it is close, coming to a minimum value of 13.1 at around 40 GPa. As
with the T = 630 K HI data, this is likely an underestimation of the true f occupation at
low temperature, because the DMFT(HI) method overestimates the tendency for a hole to
localize in the f -orbitals. Nevertheless, these results are at least qualitatively consistent
with experimental temperature results. The strong temperature dependence of nf in both
DMFT(QMC) and DMFT(HI) is consistent with measurements of Yb valence done in Yb
compounds, such as YbAgCu4 [144], YbInCu4 [145, 146]. Remarkably, in YbGaGe [147],
and YbAl3 [148], the measured valence of Yb is temperature independent. The charac-
teristics in the electronic structure of those materials that causes the Yb valence to be
temperature independent are not understood.
8.4.1 QMC Ergodicity
In the DMFT(QMC) calculations, we encountered significant difficulties with ergodic be-
havior in the Hubbard-Stratonovich field configurations. Such behavior has been seen in
the past by Scalettar, Noak and Singh [149] in the Hubbard model with U >> W where
the Monte Carlo sampling became ‘stuck’ around configurations where there was a net
spin moment, when in fact the average spin moment should be zero. This was solved by
introducing global moves which flip all the Hubbard-Stratonovich fields at once.
Shown in Fig. 8.5 are the normalized sampling distributions of configurations as a
CHAPTER 8. DMFT APPLIED TO YB VALENCE TRANSITION 139
13 13.2 13.4 13.6 13.8 14n
f
0
10
20
30
40
Prob
abili
ty D
istr
ibut
ion
4.053.903.753.603.453.353.20
Figure 8.5: Plot of probability distribution of sampling Hubbard-Stratonovich field con-figurations in DMFT(QMC) of a given f -electron count nf . Different curves correspondto different volumes, from rws = 3.10 at the bottom up to rws = 4.05 at the top.
function of nf . Generally speaking, there are peaks near nf = 14 and nf = 13, and the
relative ratios of their heights changes as a function of volume. The volumes which are
most difficult to converge are rws = 3.45 and rws = 3.60, where there is significant weight
in both the peaks around nf equal to 13 and 14. At rws ≤ 3.35, convergence is still difficult
because the distributions still maintain a broad feature that extends almost up to nf = 14.
At rws ≥ 3.90, convergence of the calculation is no problem as the distributions in this
region are almost entirely centered around the nf = 14 area with very little weight near
nf = 13.
The ergodicity problem here can’t be easily solved by global moves, as it is nontrivial
to derive a move for the Hubbard-Stratonovich fields which changes the occupation of a
single orbital from 0 to 1, or vice-versa. Previous applications of global moves have been
restricted to swapping the occupations of two orbitals, and the perscription for changing the
Hubbard-Stratonovich fields is straightforward in that case. These results show another
clear difference from ergodicity issues observed by other groups. In particular, at the
CHAPTER 8. DMFT APPLIED TO YB VALENCE TRANSITION 140
-10 -8 -6 -4 -2 0 2 4 6
ω (eV)
0
10
20
30
40
50
60
70
80
A(ω
) (e
V-1
)
4.053.903.753.603.453.303.203.10Valence
Figure 8.6: Plot of Yb spectral function A(ω) at T = 630 K calculated in LDA+DMFT(HI)for selected rws values, relative to the chemical potential. Spin-orbit split peaks correspond-ing to the f14spd2 state are near the chemical potential, while peaks corresponding to thef13spd3 state are around −U(rws).
volumes where the ergodicity is most apparent, there is a non-vanishing sampling of states
between the peaks at nf = 13 and nf = 14. This is an indication that the barrier to
transitioning between integer occupations is not large, and perhaps a brute-force approach
might be most applicable. This property is likely a consequence of Yb being a metal at
all volumes, due to weak hybridization of the overlapping metallic spd orbitals with the f
orbitals.
8.4.2 Spectral Function
Calculations of the spectral function A(ω) are presented in Fig. 8.6 for DMFT(HI). At
low pressure, there is a clear two peak structure near the Fermi energy which is attributed
to spin-orbit splitting of the f -states so that there is one each of j = 5/2 and j = 7/2
peaks. As the pressure is increased, the spectral weight of these f14 peaks decreases,
and the f13 peaks appear below at about −.6Ry ≈ U . Also, the upper peak (j = 7/2)
becomes pinned to the Fermi energy. When the valence transition is fairly complete, the
CHAPTER 8. DMFT APPLIED TO YB VALENCE TRANSITION 141
j = 7/2 peak above the Fermi energy can be interpreted as the upper Hubbard band
for the 4f states. From this we see how the Hubbard bands evolve with pressure; they
don’t gradually split (as with increasing U) but the lower Hubbard bands (f13) gradually
appear and the upper Hubbard bands gradually disappear and simultaneously move across
the Fermi energy until there is only a single j = 7/2 state left above εF . This effect is
clearly something that cannot be captured by a single-particle band theory such as LDA
or LDA+U, because they can only produce bands which have integral weight. It is not
impossible to get an f13 state in LDA+U; it can be done in certain chemical environments
[61], but it is rather difficult to achieve for elemental Yb. The mixed valent state is also
rather inaccessible for LDA+U, because LDA+U cannot create a Hubbard band for the
f -states which has fractional occupation (which would be required to capture the gradual
nature of this transition) without strong hybridization, which would not be reasonable for
narrow band f states.
8.5 Conclusion
We have examined the valence transition and equation of state of elemental fcc Yb using
LDA, GGA, LDA+U, and all-electron LDA+DMFT with impurity solvers that are either
highly accurate or very computationally efficient. We have shown how and why LDA+U
results can be suspect for a system containing Yb, particularly if Yb may be in a mixed
valent state, and that LDA+DMFT corrects these deficiencies. LDA+DMFT(HI) provides
good qualitative agreement with experiment, showing a full valence transiton to f13 over 35
GPa and the transition is gradual as seen in experiment. LDA+DMFT(QMC) improves
on this by providing excellent quantitative agreement with experiment. We examined
the temperature dependence of this valence transition, finding that both DMFT(HI) and
DMFT(QMC) are consistent with experimental results for several Yb compounds over all
temperature ranges studied.
CHAPTER 8. DMFT APPLIED TO YB VALENCE TRANSITION 142
8.6 Acknowledgements
I would like to thank Simone Chiesa, Jan Kunes, and Richard T. Scalletar for stimulating
duscussions on DMFT, QMC and Yb.
143
Chapter 9
Orbital Ordering in RbO2
9.1 Introduction
Correlated materials have generated a lot of interest recently in the literature. Typically
these materials contained localized d or f electrons which are not accurately treated in
density functional theory by the L(S)DA or GGA exchange-correlation approximations,
although there are cases where p orbitals behave in correlated manners, and can show
magnetic effects. SrN was recently predicted to show magnetism in N2 dimers [150], and
antiferromagnetism in RbO2 has been known for at least 30 years [151].
Proper treatment of these materials requires methods which go beyond local approx-
imations. Two commonly used methods are LDA+U and LDA+DMFT, which combine
LDA’s accurate treatment of the valence electrons, with a many-body orbital dependent
treatment of the correlated subset of orbitals.
Since the identification of electrons in a solid as belonging to a particular atom or its s,
p, d, etc. states is somewhat ambiguous, any treatment of correlations is going to somewhat
depend on how these states are projected out of the Bloch states. This is typically not
problematic, and most applications of both LDA+U and LDA+DMFT work quite well
with an atomic orbital projection which may be different in different codes. At least, these
methods don’t typically fail to adequately describe a system of interest because of how the
CHAPTER 9. ORBITAL ORDERING IN RBO2 144
projections are done. There are exceptions however. Generally, one should project states
onto Wannier functions and use these as the orthonormal basis for a correlated method,
and there are cases where the Wannier functions that need to be used are significantly
different from atomic orbitals. RbO2 is an interesting example because its valence bands
are composed of O p states, and the O atoms are paired up in molecular units, so it is
more sensible to think of this crystal as being a distorted rock salt structure with Rb+
as the cation, and the hyperoxide ion O−2 as the anion. An application of a LDA+U or
LDA+DMFT to this material must use projections on O2 molecular orbitals, rather than
atomic orbitals which are not an acceptable approximation to Wannier functions.
The body-centered tetragonal (BCT) structure of RbO2 is shown in Fig. 9.1. This is
actually only the symmetrized structure. Below 194 K there are several phase transitions
where the structure undergoes minor distortions from the BCT structure [152]. In this
chapter we examine the ordering of the 2p antibonding π∗x and π∗y orbitals and how this
ordering is frustrated, which is likely the cause of the complexity of the low T phase
diagram.
9.2 Tight Binding Hamiltonian
DFT-LDA calculations were performed using the abinit code [103] and the FPLO code
[22]. Abinit was used to do structural relaxation at each volume, using the PBE GGA
functional [2], and FPLO with PZ LDA [3] was used with the relaxed parameters to
produce band structures for tight-binding fits.
The LDA spin-unpolarized band structure and density of states for the experimental
P = 0 volume V0 are shown in Fig. 9.2. The conduction band character is composed
primarily of the two O2 π∗ orbitals, and other bands are well separated. From a band
structure point of view, RbO2 is definitely a molecular solid. Spin-polarized calculations
show a half-metallic ferromagnet, however this is in complete disagreement with experi-
ment, where the material is insulating and antiferromagnetic with TN = 20K. Supercell
CHAPTER 9. ORBITAL ORDERING IN RBO2 145
Figure 9.1: Structure of RbO2. Red atoms are O, blue are Rb. The structure is bodycentered tetragonal.
calculations have been reported to reveal the antiferromagnetic nature of RbO2 [153], how-
ever they are still metallic. At a compression of V = 0.65V0, shown in Fig. 9.3, the π∗
bands have broadened but their dispersion is not significantly altered. Notably, at this
higher pressure the gap above the π∗ bands has actually increased, and the π and σ bands
below the conduction bands have become more dispersive. At compressions somewhat
above this, the π∗ bands begin to overlap the π bands and the tight-binding fit does not
agree quite so well.
A schematic for the hopping channels used in the tight-binding model is shown in Fig.
9.4. To investigate orbital ordering, we construct the Hamiltonian in a cell whose base
is√
2a ×√
2a. The lattice vectors for this cell are monoclinic, being (a, a, 0), (a,−a, 0),
(a/2, a/2, c/2). For convenience, we define the following k-dependent tight-binding quan-
CHAPTER 9. ORBITAL ORDERING IN RBO2 146
Γ X P Γ3 Z Γ Γ1 Γ2-8
-6
-4
-2
0
2
4
6
0 5 10
O2
π*
O2
π
O2
σ
Figure 9.2: Band structure and density of states from spin unpolarized LDA calculation atV = V0, the experimental P = 0 volume. Conduction electrons are composed of molecularO2 anti-bonding π∗ orbitals. The unoccupied bands above the gap are mostly O2 σ
∗ andRb 5s and 4d character. The tight binding fit of Table 9.1 is shown in dashed red, withexcellent agreement.
Figure 9.3: Band structure and density of states from spin unpolarized LDA calculationat V = 0.65V0. Comparing with Fig. 9.2, bands are broadened, and the gap above theconduction band is increased. The tight binding fit of Table 9.1 is shown in dashed red,again with excellent agreement. The O2 σ bands have gained considerable dispersion inthe z direction, as seen along the Γ-Z line.
The 4 × 4 Hamiltonian for the supercell is then
H =
εx +H11xx +H+
z H11xy +H+
xy H12xx +H−
z H−xy
H11xy +H+
xy εy +H11xx +H+
z H−xy H12
yy +H−z
H12xx +H−
z H−xy εx +H11
xx +H+z H11
xy +H+xy
H−xy H12
yy +H−z H11
xy +H+xy εy +H11
xx +H+z
(9.7)
with εx and εy as the energy levels of the π∗x and π∗y orbitals (which are equivalent for the
purposes of this discussion). The k-dependence of the tight-binding quantities has been
suppressed for clarity. The 2 × 2 Hamiltonian for the primitive unit cell Hp is
Hp =
εx +H11xx +H+
z +H12xx +H−
z H11xy +H−
xy +H+xy
H11xy +H−
xy +H+xy εy +H11
xx +H+z +H12
xx +H−z
. (9.8)
CHAPTER 9. ORBITAL ORDERING IN RBO2 148
xy
2
1
tπ
tσ
t′
t′xy
tzxy
tz
Figure 9.4: Schematic of tight-binding hoppings, looking down on the x-y plane. Blackand light blue atoms are O2 molecular centers, with the light blue atoms being positionedin planes at a distance of c/2 above and below the plane of black atoms. The orbitalspictured are π∗x and π∗y orbitals, which from top-down appear as p-orbitals. The hopping
t′z (not pictured) is in the z direction, of a distance c. The base of the√
2 ×√
2 unit celldescribed in the text is shown with the dotted line, and the two atoms in the unit cell aremarked 1 and 2.
The optimized cell parameters and hoppings are shown in Table 9.1, and the hoppings
are plotted in Fig. 9.5. The bond length in the O2 molecule is 1.21 A, and in the peroxide
ion O2−2 (in hydrogen peroxide) 1.49 A[154], and the O2 bond length in RbO2 is 1.35 A
at ambient pressure, giving the confirmation of the charge state of the O2 ion as -1. The
largest hopping is tzxy, which represents hopping between layers and from π∗x to π∗y orbitals.
There are 8 neighbors which participate in this hopping, compared to 2 each for tσ and tπ,
thus since 8|tzxy| >> 2|tσ | + 2|tπ|, this would suggest that the strongest AFM coupling is
between layers. It is not known what type of AFM ordering actually takes place in RbO2.
Table 9.1: Optimized cell parameters and tight binding coefficients. Volume is expressedas fraction of the experimental P = 0 volume. Units for a and the O2 bond length are A.Units for tight-binding parameters are meV.
9.3 Poor Man’s LDA+U method
Since standard electronic structure codes use atomic projections for correlations, they
cannot be used to perform LDA+U using the molecular orbitals of RbO2. Given the
tight-binding fit of the previous section, we can construct a poor man’s LDA+U method
which acts on the Wannier functions we have constructed. We construct a new energy
functional as
Etot =∑
kn
εkn +EU (9.9)
EU = U∑
i,α
niα↑niα↓ + U ′∑
iσσ′
nixσniyσ′ (9.10)
where i is a site index and α = x, y is an orbital index. U and U ′ are Hubbard parameters
introduced here. The physical regime is expected to be where U > U ′ at all volumes. The
orbital potential obtained from ∂Etot/∂niασ is then
Viασ = Uniα,−σ + U ′∑
σ
ni,−α,σ (9.11)
CHAPTER 9. ORBITAL ORDERING IN RBO2 150
0.6 0.7 0.8 0.9 1 1.1
V / V0
0
50
100
150
Tig
ht B
indi
ng P
aram
eter
s (m
eV)
-tσ-tπt’tz
tz’
txy
tz
xy
Figure 9.5: Plot of tight-binding parameters vs. volume. Note that the magnitude oftσ and tπ are plotted. The largest parameter at all volumes is tzxy, the nearest neighborhopping between π∗x and π∗y orbitals. The change in trend at V < 0.65V0 is the resultof the interaction between the π∗ and π bands as the gap between them closes, and thetight-binding fit becoming less accurate.
where −α is used to denote ‘the other orbital’ on site i and σ denotes the spin projection.
One then takes the niα,−σ as mean-field variables, and iterates the eigenvectors of the
Hamiltonian H = HTB + V to convergence, recomputing the n values after each iteration
to reconstruct the potential matrix for the next iteration. Two considerations must be
taken into account in practice, (1) double counting of the potential, and (2) the tendency
to reach a local minimum rather than the desired absolute minimum. The potential ends up
being double counted because the application of the potential V changes the eigenvalues,
so one must subtact Tr(V n) = 2EU from (9.9). To properly find a global minimum, one
must converge this method several times, using different starting guesses for the orbital
occupations, and then compare the total energies that come out of each result to see which
is lowest.
The general trend for U in 4f and 3d atomic orbitals is for U to get smaller as l
gets smaller. A typical value of U for 3d orbitals is 5-6 eV, so a reasonable U for 2p
CHAPTER 9. ORBITAL ORDERING IN RBO2 151
orbitals might be around 2-3 eV. Molecular π∗ orbitals are not as localized as atomic
p orbitals, so a reasonable value of U would be around 1-2 eV. Since the bandwidth of
the π∗ orbitals in RbO2 is about 1 eV, it is reasonable to expect correlated behavior
here. The phase diagrams for this LDA+U method for U and U ′ from 0 to 2 eV at three
selected volumes are shown in Figs. 9.6. Two different kinds of orbital ordering were
actually found, they are termed XY and XX ordering. XY ordering is composed of the
holes in each plane occupying orbitals π∗x on one atom, and π∗y on all neighboring atoms
(in the plane). This minimizes the Coulomb repulsion between orbitals in the plane, but
frustrates the interaction between planes (the intraplane interaction is between the black
and light blue atoms in Fig. 9.4). The XX ordering is where the holes occupy π∗x orbitals
on every O2 molecule. (YY ordering would be equivalent to XX ordering, so no distinction
is made between them.) This eliminates the frustration between the planes at the cost of
frustrating the orbitals in the plane. This ordering should not be too surprising, because
the the nearest neighbors of each O2 molecule are the ones above and below in separate
planes. There are eight nearest neighbors, compared to four second neighbors. So the cost
for unfrustrating these eight bonds will be comparable to the cost of frustrating two of
the four second neighbor bonds, and with the right parameter choices the XX ordering
will show up. One must be careful about overly trusting the XX/XY distinction. Along
the line U = U ′, the two orderings are degenerate, and away from this line the energy
difference between these two phases is on the order of 1 meV = 12 K. This is coincident
with the temperature scale of the several minor structural phase changes that occur at
low temperatures. It is likely that these phase changes are occuring as a result of this
frustration.
Since the physically realistic parameters are estimated to be 1-2 eV with U > U ′, this
model predicts that RbO2 will be an antiferromagnetic insulator with XX ordering. As
volume is decreased, the bandwidth W increases and one might also expect U and U ′ to
decrease as the Wannier functions become less localized. Thus, as pressure is applied RbO2
is expected to go first through a metal-insulator transition to become metallic and then
CHAPTER 9. ORBITAL ORDERING IN RBO2 152
a magnetic transition from AFM to FM ordering might occur. Whether RbO2 becomes
a half-metal or not will be somewhat sensitive to the parameters U and U ′, and the
possibility of the magnetic transition being from AFM to PM is not excluded.
Shown in Fig. 9.7 are several plots of the density of states for the poor man’s LDA+U
method for selected parameters. Fig. 9.7a shows the DOS of an XY ordered antiferromag-
net. The unoccupied spin-up states are the π∗x orbital on one molecule, and the unoccupied
spin-down states are the π∗y orbital on the other. In this phase the magnetic ordering is
intimately tied to the orbital ordering. This is true for the XX antiferromagnet, shown
in Fig. 9.7b. In the XX AFM phase, the π∗y orbitals form normal band states, but the
π∗x orbitals split into Hubbard bands. There is a similar effect in the XY AFM phase,
however there a particular pair of π∗x and π∗y orbitals form band-like states, and the other
pair forms Hubbard bands. The ‘beginning’ of this behavior is seen in Fig. 9.7e, where the
U and U ′ parameters are large enough for orbital ordering, but not large enough to create
an insulating state. Here the π∗x and π∗y orbitals that are partially occupied split in a man-
ner consistent with creating an XY insulator, so one can imagine that as the parameters
are turned up the DOS will smoothly evolve into that of Fig. 9.7a. The beginnings of
AFM ordering are seen too, as the π∗x and π∗y orbitals are a little bit spin split in opposite
directions. Fig. 9.7f shows how this looks once AFM ordering has set in, but the system
is not yet insulating. There is clear orbital ordering as well as antiferromagnetism, but no
gap.
The graphs in Fig. 9.7 that have not been discussed so far all show metallic solutions.
Fig. 9.7c shows the half-metallic phase that spin-polarized LDA produces. In this region,
the Hubbard U is essentially acting as the effective Stoner I of LSDA. Fig. 9.7d shows
the FM metal, just before half-metallic state sets in. Fig. 9.7g shows the ‘orphaned’ XX
ordered metallic phase seen in the lower right of Fig. 9.6a. This phase is distinct from both
its XY-ordered neighbors in Figs. 9.7e and 9.7f, and it’s also distinct from its XX-ordered
cousin shown in Fig. 9.7b. The XX insulator has π∗x states above the gap, and π∗y states
below the gap, but for the XX metal all states near the Fermi energy are π∗x states.
CHAPTER 9. ORBITAL ORDERING IN RBO2 153
9.4 Conclusion
We have examined the ground-state U -U ′ phase diagram of LDA+U applied to an effective
tight-binding Hamiltonian for RbO2. Several magnetic and orbital-ordered phases are
seen. Two kinds of orbital ordering are seen, depending on which of U and U ′ is larger.
Since the physical regime is where U > U ′, this model predicts that XX ordering will
be dominant over XY ordering. These two orderings are nearly degenerate, reflecting the
frustration of the orbital ordering. AFM magnetic ordering is seen at reasonable values of
the parameters.
CHAPTER 9. ORBITAL ORDERING IN RBO2 154
InsulatorAFM XX
MetalXY
AFM XYInsulator
XY AFM Metal
XX Metal
MetalFM
Metal
Half−MetalFM
1 200
1
2(a)
U
U ′
InsulatorAFM XX
AFM XYInsulator
Metal
FMMetal
XY AFM Metal
XY Metal
1 2
1
2
00
(b)
Half−MetalFM
U
U ′
InsulatorAFM XX
AFM XYInsulator
Metal
FMMetal
XY AFM Metal
1 2
2
1
00
XY Metal
(c)
Half−MetalFM
x
U
U ′
Figure 9.6: U -U ′ phase diagram produced by the LDA+U method described in text forvolumes (a) 1.1V0 (b) V0 and (c) 0.65V0. The dashed line indicates where the gap is 0.1eV, where the gap increases in the upper right direction. The unmarked phase to the leftof the dashed line and to the right of the FM half-metallic phase in an XX ordered metal,at all volumes.
CHAPTER 9. ORBITAL ORDERING IN RBO2 155
-1.5 -1 -0.5 0 0.5
ε − εF (eV)
-6
-4
-2
0
2
4
6
Den
sity
of
Stat
es (
stat
es/e
V)
Total DOS
π*
x1 DOS
π*
y1 DOS
π*
x2 DOS
π*
y2 DOS
-1.5 -1 -0.5 0 0.5
-5
0
5
-1.5 -1 -0.5 0 0.5
-10
-5
0
5
10
-1.5 -1 -0.5 0 0.5
-10
0
10
-1.5 -1 -0.5 0 0.5
-5
0
5
-1.5 -1 -0.5 0 0.5
-6
-4
-2
0
2
4
6
-1.5 -1 -0.5 0 0.5
-5
0
5
(a) (b) (c) (d)Spin up
Spin downU = 1.0U’ = 1.5
U = 1.5U’ = 1.0
U = 1.5U’ = 0.25
U = 0.25U’ = 1.25
(e) (f) (g)
U = 0.25U’ = 0.75
U = 0.3U’ = 1.5
U = 0.5U’ = 0.25
Figure 9.7: Full and projected density of states for selected values of U and U ′, measuredrelative to the Fermi energy at V = 1.1V0. Orbitals on atom 1 are colored solid lines,and orbitals on atom 2 are colored dashed lines. (a) is an XY ordered AFM insulator.The unoccupied states are composed of π∗x orbitals one one O2 molecule, and π∗y orbitalson the other O2 molecule. (b) shows an XX ordered AFM insulator (or small gap semi-conductor). The unoccupied states are the x orbitals on each O2 molecule. (c) shows aferromagnetic half-metal, which is the same solution as given by spin-polarized LDA. Allorbital occupations are equal. (d) shows a ferromagnetic metal. (e) shows an XY orderedmetal. The unoccupied states are the π∗x orbital on atom 1, and the π∗y on atom 2. (f)shows an XY ordered AFM metal. (g) shows an XX ordered metal. The unoccupied statesare composed entirely of π∗x orbitals on both atoms.
156
Bibliography
[1] W. Kohn and L. J. Sham. Self-consistent equations including exchange and correla-
tion effects. Phys. Rev., 140:A1133, 1965.
[2] John P. Perdew, Kieron Burke, and Matthias Ernzerhof. Generalized gradient ap-
proximation made simple. Phys. Rev. Lett., 77(18):3865–3868, Oct 1996.
[3] J. P. Perdew and Alex Zunger. Self-interaction correction to density-functional ap-
proximations for many-electron systems. Phys. Rev. B, 23(10):5048–5079, May 1981.
[4] S. Yu. Savrasov. Linear response calculations of lattice dynamics using muffin-tin
basis sets. Phys. Rev. Lett., 69(19):2819–2822, Nov 1992.
[5] Xavier Gonze, Douglas C. Allan, and Michael P. Teter. Dielectric tensor, effec-
tive charges, and phonons in α-quartz by variational density-functional perturbation
theory. Phys. Rev. Lett., 68(24):3603–3606, Jun 1992.
[6] M. H. Hettler, M. Mukherjee, M. Jarrell, and H. R. Krishnamurthy. Dynamical
cluster approximation: Nonlocal dynamics of correlated electron systems. Phys.
Rev. B, 61(19):12739–12756, May 2000.
[7] Edward Teller. On the stability of molecules in the Thomas-Fermi theory. Rev. Mod.
Phys., 34(4):627–631, Oct 1962.
[8] Richard M. Martin. Electronic Structure. The Press Syndicate of the University of
Cambridge, 2004.
BIBLIOGRAPHY 157
[9] P. Hohenberg and W. Kohn. Inhomogeneous electron gas. Phys. Rev., 136(3B):B864–
B871, Nov 1964.
[10] D. M. Ceperley and B. J. Alder. Ground state of the electron gas by a stochastic
method. Phys. Rev. Lett., 45(7):566–569, Aug 1980.
[11] S. H. Vosko, L. Wilk, and M. Nusair. Accurate spin-dependent electron liquid corre-
lation energies for local spin density calculations: a critical analysis. Can. J. Phys.,
58:1200, 1980.
[12] S. Goedecker, M. Teter, and J. Hutter. Separable dual-space Gaussian pseudopoten-
tials. Phys. Rev. B, 54(3):1703–1710, Jul 1996.
[13] John P. Perdew and Yue Wang. Accurate and simple representation of the electron-
gas correlation energy. Phys. Rev. B, 45(23):13244, 1992.
[14] Yingkai Zhang and Weitao Yang. Comment on Generalized Gradient Approximation
Made Simple. Phys. Rev. Lett., 80(4):890, Jan 1998.
[15] John P. Perdew, Adrienn Ruzsinszky, Gabor I. Csonka, Oleg A. Vydrov, Gustavo E.
Scuseria, Lucian A. Constantin, Xiaolan Zhou, and Kieron Burke. Restoring the
density-gradient expansion for exchange in solids and surfaces. Phys. Rev. Lett.,
100(13):136406, 2008.
[16] Xavier Gonze. First-principles responses of solids to atomic displacements and ho-
mogeneous electric fields: Implementation of a conjugate-gradient algorithm. Phys.
Rev. B, 55(16):10337–10354, 1997.
[17] Xavier Gonze and Changyol Lee. Dynamical matrices, Born effective charges, di-
electric permittivity tensors, and interatomic force constants from density-functional